Enhancing perceptual data using large language models in environment reconstruction systems and applications

Abstract

The invention discloses enhancement of perceptual data using a large language model in an environment reconstruction system and application, and specifically discloses a method of generating a tokenized description of a physical environment using a language model. In at least one embodiment, sensor and / or observation data of an environment may be obtained and used to generate a perceptual data set. The perceptual data and approximate location data within the environment may be analyzed to identify an aligned map data set. The aligned map data and perception data may be provided as input to a trained language model, which may be trained as associated and / or fused information to generate a single consistent representation of the environment. The language model may output a tokenized description of the environment, which may be a compact but robust textual description of the environment in a language of the particular field.

Description

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] This application claims priority to and the benefit of U.S. Provisional Patent Application No. 63 / 521,627, filed on June 16, 2023, entitled “USING LANGUAGE MODELS FOR MAPPING IN AUTONOMOUS SYSTEMS AND APPLICATIONS,” the entire contents of which are incorporated herein in their entirety for all purposes. Background Art

[0003] There are various operations that may require the generation or reconstruction of realistic digital and / or virtual environments that conform to the rules and constraints of the real world, such as autonomous or semi-autonomous navigation, and robot simulation. For example, maps (such as high-definition (HD) maps, standard definition (SD) maps, navigation maps, etc.) are widely relied on for semi-autonomous and autonomous operations. Autonomous or semi-autonomous vehicles and machines may rely on these maps and real-time sensor data for navigation, positioning, path or route planning and / or other operations. For example, for systems such as those responsible for autonomous vehicle operations, perception modules can work with mapping modules to allow such maps to be aligned with what the perception model can infer from sensor data in real time. This alignment first requires obtaining an accurate self-vehicle position through a mechanism such as a global positioning system (GPS) to determine the relevant local map data, and then associating the local map data with the perception data. This approach is usually very sensitive to the accuracy and availability of location data, and places strict requirements on map quality and freshness. The separation of perception modules and mapping modules also prevents the overall solution from being used to resolve the inconsistency between perception results and map data. Furthermore, when no map is available, the perception module must be responsible for detecting all driving-related objects around the ego car in real time and with high accuracy, even though many driving-related objects (such as lanes, signs, or curbs) are static and can be drawn offline and retrieved on demand. This places higher demands on the perception module in terms of both computing resources and performance, including aspects such as accuracy and latency. BRIEF DESCRIPTION OF THE DRAWINGS

[0004] Various embodiments according to the present disclosure will be described with reference to the accompanying drawings, in which:

[0005] Figure 1 illustrates example data generated during stages of an environment reconstruction process according to at least one embodiment;

[0006] Figure 2A and Figure 2BAn example environment reconstruction pipeline and example tokenized descriptions generated using such a pipeline are shown in accordance with at least one embodiment;

[0007] Figure 3 An example pipeline for generating a tokenized description of at least a portion of an environment by fusing, in part, perception data and map data, according to at least one embodiment is shown;

[0008] Figure 4A and Figure 4B An example process for generating a tokenized description using fused sensory data and map data is shown in accordance with at least one embodiment;

[0009] Figure 5A An example map graphic is shown in accordance with at least one embodiment;

[0010] Figure 5B An example landmark analysis system is shown in accordance with at least one embodiment;

[0011] Figure 5C shows an example tokenized text string in accordance with at least one embodiment;

[0012] Figure 5D An example lane diagram is shown according to at least one embodiment;

[0013] Figure 5E An example architecture for determining output status according to at least one embodiment is shown;

[0014] Fig. 5F showing an example image of an intersection in an example map according to at least one embodiment;

[0015] Figure 5G An example process for generating a text string representation of an environment in accordance with at least one embodiment is shown;

[0016] Figure 5H An example process for generating a tokenized text string representation of a physical environment in accordance with at least one embodiment is shown;

[0017] Figure 6 Components of a distributed system that may be used to generate a tokenized description of an environment according to at least one embodiment are shown;

[0018] Fig. 7A Inference and / or training logic according to at least one embodiment is shown;

[0019] Figure 7B Inference and / or training logic according to at least one embodiment is shown;

[0020] Figure 8An example data center system is shown in accordance with at least one embodiment;

[0021] Fig. 9 A computer system according to at least one embodiment is shown;

[0022] Fig.10 A computer system according to at least one embodiment is shown;

[0023] Fig.11 illustrates at least a portion of a graphics processor according to one or more embodiments;

[0024] Fig.12 illustrates at least a portion of a graphics processor according to one or more embodiments;

[0025] Fig.13 is an example data flow diagram of a high-level computing pipeline according to at least one embodiment;

[0026] Fig.14 is a system diagram of an example system for training, adapting, instantiating, and deploying machine learning models in an advanced computing pipeline in accordance with at least one embodiment;

[0027] Fig.15A and Fig. 15B A data flow diagram illustrating a process for training a machine learning model, and a client-server architecture for augmenting an annotation tool with a pre-trained annotation model, according to at least one embodiment;

[0028] Fig.16A An example of an autonomous vehicle is shown in accordance with at least one embodiment;

[0029] Fig. 16B A method for Fig.16A Examples of camera positions and fields of view for autonomous vehicles;

[0030] Fig. 16C According to at least one embodiment, Fig.16A An example system architecture for an autonomous vehicle; and

[0031] Fig.16D A method for performing a multi-threaded communication between a cloud-based server and a Fig.16A A system for communication between autonomous vehicles. DETAILED DESCRIPTION

[0032] Various embodiments will be described in the following description. For the purpose of explanation, specific configurations and details are set forth to provide a comprehensive understanding of the embodiments. However, it will be apparent to those skilled in the art that these embodiments may be practiced without the specific details. In addition, well-known features may be omitted or simplified in order not to confuse the embodiments being described.

[0033] The systems and methods described herein may be used by, but are not limited to, non-autonomous vehicles or machines, semi-autonomous or autonomous vehicles or machines (e.g., in one or more advanced driver assistance systems (ADAS), one or more in-vehicle infotainment systems, one or more emergency vehicle detection systems), manned and unmanned robots or robotic platforms, warehouse vehicles, off-road vehicles, vehicles connected to one or more trailers, spacecraft, boats, shuttles, emergency response vehicles, motorcycles, electric or motorized bicycles, aircraft, engineering vehicles, trains, submarines, remotely operated vehicles (such as drones), and / or other vehicle types. In addition, the systems and methods described herein may be used for a variety of purposes, such as, but not limited to, for machine control, machine motion, machine actuation, synthetic data generation, generative AI, model training or updating, perception, augmented reality, virtual reality, mixed reality, robotics, security and surveillance, simulation and digital twins, autonomous or semi-autonomous machine applications, deep learning, environmental simulation, data center processing, conversational AI, light transport simulation (e.g., ray tracing, path tracing, etc.), collaborative content creation for 3D assets, generative AI, cloud computing, and / or any other suitable application.

[0034] The disclosed embodiments may be included in a variety of different systems, such as automotive systems (e.g., in-vehicle infotainment systems for autonomous or semi-autonomous machines, perception systems for autonomous or semi-autonomous machines), systems implemented using robots, aerial systems, medical systems, boating systems, smart area monitoring systems, systems for performing deep learning operations, systems for performing simulation operations, systems for performing digital twin operations and systems implemented using edge devices, systems including one or more virtual machines (VMs), systems for performing synthetic data generation operations, systems implemented at least in part in a data center, systems for performing conversational AI operations, systems implementing one or more language models such as a large language model (LLM), systems for performing generative AI operations (e.g., using one or more language models), systems for performing light transport simulations, systems for performing collaborative content creation of 3D assets, systems implemented at least in part using cloud computing resources, and / or other types of systems.

[0035] Methods according to various illustrative embodiments provide for generating tokenized representations of one or more regions or domains within a physical environment. In particular, various embodiments may use a large language model (LLM) or other generative artificial intelligence (AI)-based method to generate tokenized representations of regions or domains. For example, a language model may be trained to represent regions based not only on low-level primitives that may be determined from captured sensor data, but also on aspects such as semantics, topology, and geometry associated with these primitives, and relationships between objects that may be determined using these primitives. These tokenized representations may correspond to or be used to generate feature vectors, embeddings, or points in a latent space representing each region or domain, as well as other such options. These feature vectors may be used in aggregate to represent an entire environment or set of regions.

[0036] In at least one embodiment, the accuracy of the generated tokenized description can be improved by enhancing the perception data with local map data or otherwise fusing the map data with the perception data to generate a single consistent representation. A set of observations of the environment can be obtained, wherein the observations can correspond to sensor data captured using one or more sensors of similar or different types. The observations can be analyzed by a perception module to generate a set of perception data, wherein the perception data can be related to objects identified in the environment and determined or inferred aspects of these objects. A current and / or reference position in the environment can also be determined, which can be used to identify localized map data. The localized map data can be analyzed together with the perception data to generate a set of aligned map data, wherein corresponding objects in the map data and the perception data are identified and associated, so that the position of a point in the perception data is aligned with the point of a corresponding feature in the map data. At least a portion of the aligned perception data and the map data can be provided as input to a trained language model. The language model can analyze the map data and the perception data and attempt to generate a single fused representation of the environment based at least in part on the map data and the perception data. The model can use its domain-specific learning and semantics, relations, topology, geometry, and other information provided with or determinable from the map and sensory data to reason about a consistent representation of the environment. The consistent representation can be used to generate a tokenized description of at least a portion of the environment. The tokenized description can correspond to a string of text-based tokens written in a domain-specific language. The tokenized description can be a compact and discrete representation of the environment that is lightweight enough to be processed in real time or near real time, but robust enough to include the necessary information for making decisions related to one or more target operations. In at least one embodiment, the compactness of the tokenized description can be improved by training a language model to generate tags only for those objects or related aspects of those objects that are determined to be important for a given task, operation, or domain. The ability to reason about a consistent representation from maps and sensory data allows the generation of useful tokenized descriptions even when maps and / or sensory data may be unavailable, incomplete, inaccurate, or otherwise unreliable.

[0037] In at least one embodiment, a language model can be used to generate a representation of an operational design domain (ODD) such as an intersection using a tokenized representation written in a language such as a road topology language (RTL). An embedding can be generated for each such ODD, allowing each ODD to be represented by, for example, a point in an n-dimensional latent space. Similar ODDs such as similar intersections will have similar embeddings. The generated embeddings can capture the semantics, geometry, topology, and / or other information of the ODD.

[0038] In at least one embodiment, such a language model can generate a representation of an environment that conforms to real-world rules and constructs and takes into account omissions or errors in the input data used to generate the representation. Objects in the environment can be represented using individual tokens or sequences of tokens, optionally wherein a tag descriptor provides semantics and other information associated with these tags. Text-based representations can be one-dimensional strings of these tags and tag descriptors that can encapsulate important spatial information and semantics of an environment. The advantage of such text-based descriptions is that it can be discrete and compact, allowing fast processing, searching, updating and other such operations. The text-based representation of the generated environment can be used to generate many other types of representations useful to various operations or tasks, such as a bird's-eye view map, a high-definition (HD) map or a 3D virtual environment, and other such options.

[0039] In at least one embodiment, generative AI can be used to provide semantic understanding of the environment based at least in part on sensor data of the captured environment. These sensor data can be processed and fed to a trained generative AI model (such as a large language model or "LLM"), for example, the trained generative AI model can output a text description of the environment in a structured text format (such as a road topology language (RTL)). A tokenized representation of a map or graph of the environment can be provided in a text string in RTL. Generative AI can be trained in such a way that it can fill in gaps or correct errors in sensor data based on a semantic understanding of objects or elements in the environment. The language model can receive input including semantic, positional and / or geometric information of the determined environment, such as by processing sensor data (e.g., images or LIDAR data) of the captured environment, and can update the text representation of the scene when the environment changes due to movement or other such events. The language model can also take other inputs such as previous maps or contextual information (which indicates weather, time of day, season, urban / rural area, geographic location, and the like). The input data can be represented by embeddings, feature vectors, or points in a latent space, which allows for a relatively simple search for similar environments. In this way, especially when the current environment or situation may not have enough data available to make a high-confidence decision on the action to be taken, the action to be taken in the environment can be quickly determined by determining what actions have been taken in similar environments. The ability to determine what others have done in similar environments can help the system operate in a manner similar to how humans use "intuition" in a given situation, even in the presence of possible data loss, such as when snow may obscure the route along the road, but humans can infer where to drive based on other information available in the environment. This approach can be used for a variety of geospatial information processing and autonomous driving tasks (such as map building, map editing, map-based navigation, planning and driving) by representing these tasks as document manipulation tasks. Once generative AI is trained, it can also be used to generate realistic simulated environments that conform to real-world rules, such as it can be used to test autonomous vehicles or robots. This method can also be used to correct or update noisy or partial environmental maps or maps. A generative AI model may take sensor data directly as input, or may receive input generated from sensor data in one or more stages in the pipeline, such as a stage that extracts features and generates embeddings of those features in a latent space that can be provided as input to a generative model.

[0040] Methods according to various illustrative embodiments may provide for mapping in autonomous or semi-autonomous systems and applications using language models. Systems and methods are disclosed for performing various map operations such as map building, map editing, map-based navigation, route making, planning and perception, error checking, data cleaning and data validation, and other similar operations using one or more language models (e.g., LLMs). For example, a deep learning model (such as an LLM) may encapsulate domain knowledge about how a road network and / or objects are structured. By training the LLM to predict the structure and properties of a graph described in a domain-specific language (DSL) such as RTL, the LLM learns to establish the correct relationships between objects on the road. The RTL may express roads, objects, and / or other map-related information using a language (e.g., by modeling the relationship between lane elements and other map features), such that in addition to a natural language or conversational language, the LLM learns to interpret the RTL to generate output. An automated process may be implemented to convert existing map information into RTL, and to convert the output of the LLM from RTL into a suitable map format (e.g., the format of an HD map deployed in a production vehicle). Such LLMs can be used to solve various challenging problems associated with mapping technology, such as identifying or correcting errors or gaps in maps, creating maps from photographs or video streams of road data, creating maps from aerial or satellite imagery, and / or creating maps from text descriptions. Once created, maps can be used for a variety of tasks, such as for autonomous vehicles (AVs) or autonomous systems, semi-autonomous vehicles or systems (e.g., for advanced driver assistance systems (ADAS)), simulation systems (e.g., for developing or testing / validating AV / ADAS algorithms or for creating training data for AV / ADAS perception), etc.

[0041] Variations of this and other such functionality may also be used within the scope of the various embodiments, as would be apparent to one of ordinary skill in the art based on the teachings and suggestions contained herein.

[0042] Figure 1An example data processing flow that can be implemented in an environment representation and / or reconstruction system according to at least one embodiment is shown. In this example, sensor data 104 (or other raw data captured or representing the environment) about a particular environment 102 is obtained. The environment can be any appropriate physical environment, such as an indoor or outdoor environment that can include any number of different types of objects or elements. The sensor data can include data captured or obtained using any sensor of a plurality of different types of sensors, which can include cameras, LIDAR systems, radars, sound sensors, distance sensors, etc. In various embodiments, additional data related to the environment 102 can also be obtained, such as additional data related to basic map data, context data, motion data, or other such data that can also be obtained for a virtual environment, an augmented environment, or an enhanced environment. In this example, the sensor data 104 (and any other available and useful data) can be used to generate an initial representation 106 of the environment 102. In at least one embodiment, this can include a point cloud representation of the environment 102, which is generated by analyzing and aggregating the sensor data 104 that may have been captured by multiple sensors to generate a single n-dimensional (e.g., 2D, 3D, or 4D) representation of the environment. Other initial representations may also be generated, which may depend at least in part on the type of sensor data provided. If image data is provided, the image data may be analyzed in an attempt to determine feature and depth information, which may be combined from multiple images from different viewpoints in an attempt to generate at least a 3D representation of the environment 102 or at least objects and shapes within the environment.

[0043] This initial representation 106 of the environment 102 can be analyzed to try to determine specific aspects 108 of the environment. For example, a point cloud can be analyzed to determine the categories (or types) of objects represented in the environment, such as roads, traffic signs, sidewalks, buildings, etc. The representation can also be analyzed to try to determine the locations of these objects in the environment, such as a set of 3D coordinates relative to the determined origin location. The initial representation 106 can also be analyzed to try to determine various relationships between these objects, such as where a crosswalk crosses a specific lane, or where a stop sign is associated with a specific lane and indicates an expected behavior. Once these determined aspects 108 are obtained, these aspects can be used to generate an object-based representation 110 of the environment 102. Various other types of representations can also be generated within the scope of various embodiments. As shown in the figure, in this example, the object-based representation 110 will not be a comprehensive description of the environment 102, but will focus on various types of objects or features of the environment that may be related to a specific task. For example, for autonomous driving, the object-based representation may include objects such as road lanes, crosswalks, intersections, etc., but may not include objects that may not be directly related to driving, such as buildings, billboards, mailboxes, and other such objects, unless to some extent these objects may be relevant to a particular operation or task. In this example, the object-based representation 110 also does not include vehicles, pedestrians, or other movable objects that are only at a particular location in the environment 102 at a particular time, but any or all of these and other such objects may also be included in the representation within the scope of various embodiments.

[0044] From this object-based representation, an object graph 112 can be generated that provides different representations of the environment 110. An advantage of the object graph 112 is that it is relatively lightweight and can be used to compactly describe aspects of the environment 102 that are important for a particular task or operation. For example, such an object graph 112 can be provided to a map generator to generate an HD map (or other such map or representation) that can be provided to an autonomous vehicle to make navigation decisions. Such an object graph 112 can also be provided as input to an environment generator, which can generate a realistic 3D virtual environment that can be used for tasks such as robotic simulation or digital world recreation. A large number of object graphs can be stored to represent many different environments, which can require significantly less memory or storage capacity than sensor data (e.g., a large number of high-resolution images). Such object graphs can also be quickly analyzed to allow real-time operations, such as autonomous navigation or control.

[0045] One challenge faced by existing approaches to generating such representations is the limited ability to perform automated geospatial information processing, particularly using algorithmic frameworks that are general enough to support a wide variety of use cases. Existing solutions often have task-specific designs that cannot be easily adapted to new task requirements, contain built-in assumptions that may not always hold true in real-world situations, and do not effectively leverage available data and human input. Existing approaches are also limited in their ability to learn from large amounts of diverse data related to these different tasks or use cases. Many existing solutions rely heavily on domain expertise and manually designed logic or rules at various steps in the processing pipeline. These attempted solutions are difficult to accurately complete and improve, requiring manual effort to adjust the results and get them correct. Improvements to these systems are expensive and often provide increasingly small performance gains for the effort expended.

[0046] The method according to at least one embodiment can provide a general method for processing information about such an environment 102 (such as geospatial information and semantic information). In at least one embodiment, a deep learning model that encapsulates domain-specific (or agnostic) knowledge about how objects in the environment are structured and associated can be used. An example deep learning model is a large language model (LLM) that can be trained to generate a text description of the environment, which retains a semantic understanding of the environment in addition to providing information about the categories and locations of objects in the environment. In at least one embodiment, the LLM can generate a tokenized text string as a representation of the environment, where the objects in the environment are represented as tags in the string. There can also be a set of tag descriptors in the string, which are associated with specific tags, and these tag descriptors provide semantic and / or relationship information related to each tag of the string. For example, in addition to generating a compact and sufficient representation of the environment, the advantage of using a model such as an LLM is that the LLM can fill in gaps in sensor data or make corrections when it is necessary to provide a more accurate representation of the environment. For example, training the LLM to predict the next tag in the text string (e.g., corresponding to the next object in the object graph) can help the LLM learn to establish the correct relationship between objects in the environment. For example, this can include identifying or correcting errors or gaps in the environment representation, creating an environment representation (e.g., a map or object map) from photographs or video streams of environment data, creating an environment representation from aerial or satellite imagery, and creating an environment description from textual descriptions, among other such tasks.

[0047] In at least one embodiment, a language model-based approach can be used that allows model training on large-scale existing environmental representations (such as maps), which makes data-driven performance improvements easier and more scalable in terms of domain expertise. A training method that can specifically teach LLM to recognize the next tag in a graph can be used. In at least one embodiment, LLM can generate a deep underlying representation of how objects in the environment are connected or associated, as well as a model of graph data that has been presented to LLM as input. In at least one embodiment, various tasks in geospatial information processing can be unified under a shared formula so that the same algorithm model can be reused without additional engineering effort. Processing efficiency can be further improved by replacing manual labor with automation based on machine learning models. A large language model can be trained on a large amount of environmental data, so that it can automate various tasks, such as missing element detection, inaccurate element correction, and reasoning about relationships between elements and other such tasks. Each task can be implemented without heavy reliance on human expertise for explicit design, and can be continuously improved with additional training data. Such a model can utilize existing environmental (e.g., map) data without additional data management and annotation costs. Models can be trained in a task-agnostic manner, allowing the models to be extended to other use cases without significant additional effort. These representations can include or be used to generate high-quality maps useful for tasks such as those related to advanced driver assistance systems (ADAS), autonomous vehicles (AV), unmanned aerial vehicles (UAV), or simulation systems, such as useful for developing or testing / validating AV / ADAS / UAV algorithms or creating training data for AV / ADAS / UAV perception.

[0048] The method according to at least one embodiment attempts to improve, optimize or at least control the way of perceiving the environment. In various existing systems, the perception of the environment is relatively primitive and based on the rules of detected objects. For example, an existing system may analyze the captured image to identify the position of road lanes and lane markings, but has no concept of what the lines on the road mean or how these lines are related to nearby road signs or traffic lights. Existing systems may identify these objects and use the positions of these objects to generate a map reflecting these objects. The system may try to determine the relationship and apply rules to these objects to ensure that the placement is meaningful and determine any relationship, but this is usually done during post-processing when most other data has been discarded. Applying rules based on detected objects means that if the relationship and semantic meaning of the various objects in the scene are known and used to generate the process of representing the environment, it is difficult to detect gaps, errors or omissions that can be detected in other ways. In addition, in many cases, rule-based methods are more difficult to expand.

[0049] The method according to at least one embodiment can obtain and apply this knowledge early in the process. As previously described, a large language model can obtain input related to the semantics, position and relationship between various objects in the environment, and can use this information to determine how to generate a realistic environment representation based on the input based on its learning, which can be made up for the fact that the input data may be somewhat incomplete or wrong. By representing the environment with text, the language model can apply its learned knowledge (learnings) to determine how to structure the representation to ensure authenticity and completeness, and fill in the blanks in the input data based on the knowledge learned from similar situations. The advantage of the language model is that the text is used as input, rather than other large instances of image or sensor data, and the text can be processed relatively quickly during training. This allows the use of millions or even billions of such documents to train generative models using self-supervision, which provides a better understanding of behaviors and relationships and which behaviors and / or relationships are applicable to a given environment or situation. For example, by converting an object-based or feature-based representation into a language representation, this text-based representation can be used to train a language model to understand the various correlations between object categories and their relative positions, including ways that may be difficult to fully enumerate. Trying to capture all relevant real-world correlations, relationships, and other semantic aspects would be extremely difficult using only the explicit rules required by various existing systems.

[0050] exist Figure 1 In the example of , the language model can take the object-based representation 110 as input and generate content that is essentially a tokenized text string representation of the object graph 112. In other embodiments, the language model may be able to take other inputs that will allow at least some steps in the generation pipeline to be eliminated as separate steps performed by separate processes or components. For example, the LLM can be trained to take a set of determined aspects 108 (e.g., semantic, topological, or geometric information of the environment or objects in the environment) in text format, and generate a tokenized text string representing the object graph 112 without having to generate an object-based representation. Similarly, in some embodiments, the LLM can take an internal representation 106 or even sensor data 104 as input without the need for a separate intermediate representation. For example, a model (as part of the LLM or a separate model) can analyze the sensor data 104 of the environment and encode the features of the sensor data into a latent space (or other embedding). The LLM can then take as input a feature vector that is a function of these individual latent space encodings, and can directly generate a tokenized text string representation of the environment. The extracted features can include semantic, relational, and geometric features and other such options. Encoding such features in the latent space prevents this information from being discarded early in the generation process and allows for more accurate representations or reconstructions to be generated.

[0051] In at least one embodiment, a tokenized text string may include a sequence of tokens, each of which represents an object in an environment. The string may also include a set of "token descriptors" that provide some semantic context or other useful information for a given token. The tokens may also be in a specific sequence, which not only helps to generate an object graph from a text string, but also allows semantic learning to be applied to a sequence of tokens, just as LLMs are usually done for words in a sentence. Such an environment can be represented using a variety of languages, as long as the language can provide a representation of a sequential representation as a discrete token. In at least one embodiment, a custom language including specific tokens and token descriptors that can accurately and compactly represent a specific type of environment can be used. For example, a road topology language (RTL) including terminology and syntax for representing map data of an environment including roads can be used. A unified, sequential, tokenized text representation can be used to model an object graph, and an object graph can be quickly generated from such a sequence of tokenized text strings in a consistent and repeatable manner. The language model can be trained to understand and "speak" in at least one specific language such as RTL. A trained LLM will know how to manipulate or fill in sentences in natural language, so the LLM can learn to fill in text strings with a structured representation language. The LLM can also reason about the relationships between objects based on its understanding of the language. The LLM can then generate a unified text representation of the environment, which may include information that did not previously exist or could not be determined from the input alone, but allows the environment to be more realistic and conform to the rules and / or constraints of the real world. For example, these may include local traffic rules or regulations, customs, and the capabilities of objects in the environment, as well as other such options. The language model can be trained to learn the semantics and syntax of the language, as well as the basis behind the semantics and syntax, including the physical concepts behind various object relationships. The LLM can consider lane boundaries as being associated with the lanes of the road, which are provided together with various requirements, traffic rules or behaviors, and related objects, rather than considering lane boundaries as lines in space.

[0052] A language model trained to generate representations using such a language can be used in at least one embodiment to describe the physical layout of an environment, such as for generating a high-quality map. The model can also generate text for describing other aspects of the environment, such as characters, animals, vehicles, or other objects and elements that can move or change position or posture over time and may only exist in the environment for a limited period of time. For example, a text string can be generated that provides such a representation that includes a map view showing where a vehicle can navigate and also includes representations of pedestrians, other vehicles, buildings, or other types of objects or entities that may be important for navigation or other such tasks. For example, if a language model is able to generate a representation that accurately describes various aspects of an environment, including nearby vehicles and pedestrians, such a representation can be used to make navigation decisions without the need to identify these objects separately and provide them as additional input to a navigation or control system. An example perception map or representation can be generated that can include any or everything in the environment that can be perceived using available sensor data (or other such data) and an understanding of the physical rules or relationships of such an environment.

[0053] Figure 2A An example pipeline 200 for generating a text-based representation of an environment according to at least one embodiment is shown. This method can automatically generate representations from a variety of different types of input data without requiring at least a certain amount of manual interaction. In this example, a capture device 202 may include one or more sensors 204, 206 that can capture or generate information about an environment 208, or be associated with one or more sensors 204, 206. The capture device may include any device, system, or component that can obtain sensor data from one or more sensors and process the sensor data or send the sensor data for processing, such as a desktop computer, a smart phone, a vehicle or robot component with data processing capabilities, and other such options. The sensor may include any appropriate type of sensor that can capture or generate useful information about the environment, including sensors such as cameras, infrared (IR) sensors, ultrasonic sensors, depth sensors, LIDAR systems, radar systems, or other such sensors or data capture elements. The environment 208 may include the environment in which the capture device 202 is located, or the environment within the capture distance of one or more sensors 204, 206.

[0054] In such an example, the capture device 202 may provide sensor data to be analyzed by the feature extraction module 210. As previously described, feature extraction may be performed as part of the large language model 212 or by a separate model or algorithm, among other such options. In this example, the feature extraction module 210 may include an encoder that may extract features from various instances of sensor data and encode those features into embeddings or points in a latent space 214. In at least one embodiment, the environment 208 may be represented by a set of embeddings or points in the latent space, which may then be represented by one or more feature vectors corresponding to those individual embeddings. The latent space 214 may be an n-dimensional latent space, where each environment (or state of an environment) may correspond to a point (or vector) in the n-dimensional latent space.

[0055] In this example, at least one feature vector representing a point in n-dimensional space may be provided as input to the large language model 212. Various other types of embeddings or representations may also be used within the scope of various embodiments. As previously described, in at least one embodiment, each object in the environment may be represented by a token in the text string to be generated and an embedding, feature vector, or point in the n-dimensional latent space. Such a feature vector or embedding may not only specify the type of an object, but may also represent various features of the object that may help encode, for example, semantic information, geographic information, and / or topological information of the object.

[0056] The language model can use the input to generate a tokenized text string representing the environment. In this example, the language model also receives other inputs that can help generate a more accurate representation. For example, the language model can receive a previous or partial map or environment representation, or a previous tokenized text string (e.g., for a previous time point or a nearby location), the language model can refer to the previous tokenized text string, and the previous tokenized text string can help the consistency of these representations over time (such as in the case of reconstructing the environment for a vehicle moving in the environment), and comparing the reasoning at each time point can help improve accuracy by reducing noise or removing false positives (or at least marking reasoning that is not meaningful based on previous determinations, such as when the object type has changed or suddenly appears without warning). Various other types of inputs can also be provided. For example, a user can use a client device 218, such as a desktop or laptop computer, to provide input that can guide the generation of a tokenized text string. For example, the client device can provide contextual information that helps guide the generation. For example, the contextual information may include the type of environment, such as an indication of an urban or rural environment, which can help the model apply an appropriate set of rules. For example, some relationships between road objects in downtown Manhattan may be very different from some relationships between road objects in rural Montana, but various other relationships may be very similar. The context information may indicate the state or country in which the sensor data was captured, as different states or countries typically have different traffic or behavioral rules, such as which lanes a vehicle is allowed to turn into at an intersection. The context information may include information about the weather or time of day, as sensor data for snowy, rainy, or nighttime environments may lack data about objects that might have been observed during a clear daytime capture period. Additionally, different behaviors or rules may be appropriate at night or in other situations where visibility may be limited. As discussed later herein, in the case where a simulated environment is to be generated based on embeddings in the latent space 214, for example, additional input from the client device 218 may help determine aspects of the simulated environment to be generated.

[0057] The tokenized text strings generated by the language model 212 can be provided to various components for various tasks. In some embodiments, the reconstruction of the environment 208 can be performed by a reconstruction module 216 or system, such as to generate a high-definition map or 3D digital model of the environment 208. In some embodiments, the text strings and / or reconstruction can be provided to the control or navigation system of an autonomous vehicle or robot to allow decisions to be made about how to move or interact with objects in the environment. In this example, the initial capture device 202 can be on or part of a vehicle, or in some embodiments can be the vehicle (or robot, etc.) itself. The reconstruction of the environment can be provided back to the capture device for performing a specific task. For example, if the capture device is an autonomous vehicle or a driver assistance system, the reconstruction (or in some embodiments, the tokenized text strings) can be provided back to the capture device to perform operations based in part on the reconstruction, such as making navigation or operating decisions, and the capture device uses associated sensors 204, 206 to capture initial sensor data.

[0058] In at least one embodiment, the reconstruction may be provided to a client device 218 for presentation or analysis, which may be the same client device that instructed the reconstruction. In some embodiments, the client device 218 may analyze the accuracy and completeness of the reconstructed environment, or may perform various operations or simulations on the environment. The client device 218 may also provide additional information such as context to the reconstruction module for generating the environment. For example, the client device may instruct the reconstruction module 216 to generate multiple reconstructions of the same environment 208 using the same tokenized text string but under different conditions. This may include, for example, versions of the same environment during the summer day, winter night, Europe and Asia (which may affect the language and style used). During model training, the tokenized text string and / or environment reconstruction may be compared with appropriate truth data to determine a loss value and update the parameters of the appropriate model.

[0059] In this example, the feature extraction and language generation operations can be part of the same model or separate models. For example, a first model (e.g., an encoder) can take sensor data as input and output a set of embeddings or latent feature vectors as output, which are then provided as input to a generative model (e.g., a large language model). In another embodiment, the generative model can include feature extraction or analysis capabilities and can generate a tokenized text string as output without any intermediate or other steps to process or analyze the input sensor data. Back to Reference Figure 1, a language model can be trained to take input from any of the various stages of the representation generation pipeline. For example, a language model can take raw sensor data as input, or an initial representation (e.g., a point cloud) generated by analyzing the sensor data using a separate module, system, component, model, algorithm, or process can be taken as input. Similarly, the model can take determined aspects or information that can be related to the semantics, topology, or geometry of the environment, or can take object-based representations generated for the environment as input and other such options. In at least some embodiments, the type of input to be used may depend at least in part on the system in which the language model will be used, because different systems may already provide specific outputs to be used. In at least one embodiment, a language model can take raw sensor data and such intermediate representations as input in order to try to provide a more accurate or more consistent representation. In some embodiments, multiple language models can be used. For example, a language model can be used to determine the semantics, topology, and geometry of the environment, which is then provided as input to another language model.

[0060] In some cases, the tokenized text string and / or reconstruction of the generated environment may be incomplete due to any of a number of reasons. For example, it may be snowing and some lane markings may not be visible or accurately represented in the sensor data. In other cases, a stop sign may be damaged and not visible, or there may be objects obscuring portions of the environment. In such cases, it may be difficult to accurately or successfully complete an action or task based on the lack of complete information. However, using methods such as Figure 2AThe system shown can make decisions based on other decisions made in similar situations and environments. For example, during the feature extraction and encoding phase, points can be determined in an n-dimensional latent space representing the extracted features. Although the set of features will be incomplete due to incomplete sensor data, a search for nearby points can be performed in the latent space 214, which will represent very similar environments. For example, this can include similar intersections at different locations, or even embeddings previously generated for the same general environment. Other types of feature-based or embedding-based searches can also be performed within the scope of various embodiments. Once similar environments are determined through such searches, a database 220 of actions taken in these environments can be analyzed to try to determine the actions to be taken in the current environment 208 with at least the minimum probability or confidence. For example, lane markings or traffic signals may not be visible, but based on what other vehicles (even the same vehicle) in other environments do in similar situations, the correct actions to be taken in the current environment 208 can be determined. This approach can allow the vehicle navigation system to function similar to human intuition, where the most appropriate action to be taken can be decided based on available information and information that worked correctly in similar situations in the past. For example, this approach could allow an autonomous vehicle to operate in the snow even when lane markings are not visible based on the location of objects such as road signs, traffic signals, other vehicles, etc., where the appropriate location and actions for the vehicle can be determined based on what other vehicles have done in similar situations. In these similar situations, lane markings may be commonly visible, so the data should be reliable. This approach could help further fill in gaps or omissions in the input data.

[0061] The search process based at least in part on features, embeddings, or tokenized text strings can be relatively fast and lightweight, which is very important for tasks such as real-time vehicle navigation or automated control. The search can also help provide additional information for generating more accurate text strings or reconstructions. For example, when a set of feature embeddings for an environment is determined, a search can be performed to identify other environments that are very similar to the current environment or the same environment at a past time point. Once determined, information about these similar environments can be analyzed to try to provide additional information, such as contextual information that can help text strings or reconstructions be more accurate. For example, the location and relative spacing of various road signs may help indicate the country or region in which the environment is located, the weather, the language in which the signs may be presented, etc. In some embodiments, embeddings or feature points from these similar environments can be provided as input to a language model in an attempt to provide more accurate tokenized text strings, or to better understand the relationship between objects detected from sensor data.

[0062] The ability to compare the current situation to similar situations in similar environments can also help with other tasks. For example, previously observed behavior of people and vehicles in similar environments can be used to predict what similar objects are likely to do in the current environment, which can help make better decisions about the best course of action to take at the current time in the current environment. For example, where and when people are allowed to cross a street and how likely they are to cross a street may vary in different locations, and knowing the likely actions at a given location based on past observations of similar environments can help avoid collisions or other undesirable events. This approach can help provide a semantic "understanding" of the environment at a specific point in time, and can help generate various types of representations, or determine various actions to take based at least in part on this understanding and knowledge of similar environments.

[0063] Reference again Figure 1 Description, the tokenized text string can effectively provide different representations of the object graph. This can allow such a method to be used in existing systems or processes that expect such a graph as input. The method proposed in this article can provide an accurate object graph representation in the form of a tokenized text string, for example, in most cases, a tokenized text string can be quickly, accurately and automatically generated without human intervention. As mentioned above, such a process can also help fill in blanks or make corrections in object graphs that may not be determined from sensor data or related inputs. In at least one embodiment, especially in the case where the language model undergoes continuous learning, the model can learn new relationships or object types that can help build a more robust object graph representation and can infer additional semantics or relationships, which can help these object graph representations become more accurate over time. The syntax of the relevant description language can be updated over time to more accurately capture or reflect these additional learnings. In at least some embodiments, the tokenized text string can be equivalent to an object graph, just in a different form. In other embodiments, the tokenized text string may include additional information that provides more context, understanding, or insights than can be obtained using a conventional object graph, and may include relationships that may not be indicated using such an object graph, including relationships that may not be easily explained using natural human language.

[0064] As previously described, the language model can apply the learned rules to the environment in a manner similar to how the language model applies language rules to natural language text. Similar to how the model learns the correct sentence structure, the model can learn the correct environment structure, such as how lanes and roads are related to each other and how they are allowed to be designed. This can prevent the language model from generating text strings indicating that lanes cross each other outside intersections, that some intersections may not have traffic signals or stop signs, that the end of the entrance ramp may not have lanes connected to the highway, etc. A semantic understanding of these relationships can help fill in this information, even if the sensor data does not include enough data to provide this information, or it is not clear how to interpret the sensor data. The language model can use its learning and semantic understanding to appropriately interpret the available data, and in at least some embodiments, it can refer to data of similar environments when appropriate or necessary. In some cases, objects that cannot be identified with a sufficient confidence level may be observed, such as objects that are partially occluded or damaged, or objects that belong to a type or style that has never been encountered before. The language model can rely on its learning to determine the type of the object more accurately and / or more confidently based on, for example, other objects in the environment and the types of objects that are generally related to these objects. For example, intersections will typically have stop signs or traffic signals, whereas highways will not and are more likely to have express lanes or mileage markings. The ability of a language model to know what types of objects to expect for a given environment and / or context, and where those types of objects are typically located in that context, can improve aspects such as object recognition, even for objects that have not been encountered before or are at least partially occluded.

[0065] In some embodiments, such as when an environment for simulation is to be generated that conforms to real-world rules, the user can enhance the sensor data to include additional (or replacement) objects in the reconstruction. For example, a user can use a client device 218 to submit information about a pedestrian bridge to be added to the environment represented by the captured sensor data. The appropriate embedding of the bridge can be determined and encoded into the latent space of the environment. In some embodiments, the user can view the reconstruction on the client device 218 and make modifications, which can be provided to the language model 212 as updated input to provide updated tokenized text strings and environment reconstruction. In some embodiments, the user can generate a new environment reconstruction independently of the sensor data. The user can provide input (e.g., speech or text) describing the environment, and the input (after any appropriate change format or analysis) can be used to select an appropriate point in the latent space 214, which can then be provided as input to the language model to generate an appropriate tokenized text string. In some embodiments, the user input can be directly provided to the language model as input without the need for separate feature extraction or embedding generation. This approach is useful for simulation environment generation, where large amounts of environment data can be synthetically generated without huge cost or manual effort, which is beneficial for training machine learning models or other artificial intelligence systems to operate in these various simulated environments. The environment generation process can then generate environments automatically, in response to human prompts, or through a combination of both. Modifications to the environment can be made relatively quickly without the extensive processing that comes with updating tokenized text strings.

[0066] In at least one embodiment, the environment generation and / or reconstruction system can work with various data formats and can appropriately perform format changes or restrictions. For example, data can be received in a map, object or graph format and can be converted into a tokenized text string in a structured language. Similarly, such a text string can be used to generate any of these or other such representations of the environment. The text can also be regenerated to correspond to different human languages, because for similar types of objects or relationships, the same language (such as RTL) can have different terms or descriptors in different human languages ​​(such as French or Spanish). When specifying contexts such as countries or regions, the language model can also learn to speak a language that it may not have been trained on initially, and can learn to use terms appropriate for a given location or context. It is possible that the components of the system all use a structured tag-based language internally, but can accept input in any of a variety of different formats or generate output in any of a format. Using a structured language for internal communication helps ensure that data about semantics, relationships or other such aspects will not be lost during processing and analysis because of the type of format used.

[0067] As previously mentioned, language-based representations can be very compact and discrete. Such aspects make language representations conducive to use in real-time, real-world environments because the representation can be updated quickly and accurately, and can be updated to include only the information that is relevant at the current time. For example, a language model can be used with the navigation system of an autonomous vehicle to make real-time navigation decisions. The ability to make these decisions is critical for many such applications. As the vehicle moves, the language representation can be updated to include, for example, parts of the environment that are now visible to the sensor in front of the vehicle, and parts that are no longer visible or are otherwise determined to be unimportant to navigation and the current position given the current direction and rate of motion can be removed or deleted, or other such aspects. Similarly, when another vehicle enters a road near the current (e.g., self) vehicle, the representation of the other vehicle can be added to the language representation, while vehicles that leave the road or are more than a threshold distance from the self vehicle can be removed. This approach can allow the language representation to be easily resized so that it can contain all information determined to be important and exclude any information that is determined to be irrelevant or at least no longer relevant based on the current position, speed, direction, etc. Keeping dynamic language representations current and compact can help make better and faster decisions by including only the information needed to make a decision at the current (or near future) point in time or period.

[0068] As previously mentioned, the language model in at least one embodiment can be self-supervised. The language model can be trained to understand the structure, pattern, syntax, relationship and other aspects of one or more languages ​​on which it is trained. Then, the trained model can obtain the text (or other language input) of the new environment, and generate or rebuild the environment based on its learning, or can obtain an incomplete or inaccurate representation of the environment, and can generate a corrected or more complete representation. This can be an end-to-end automated process without human intervention, compared to previous systems that require human intervention at some stages of map generation, even if not all stages. When the model learns that it has made a mistake and / or is able to correct errors or omissions, the model can learn from it to make better future decisions. This method can help generate more accurate representations, which is possible or at least practical for human-generated systems, because for environments such as intersections or parking lots, there may be many more rules or relationships than the actual rules or relationships that humans try to accurately encode, especially when many of these rules or relationships may be implicit, so that humans may not even consciously realize their existence. Language models can learn these and other such rules and relations without encoding or supervision, providing significant advantages over previous mapping or reconstruction systems.

[0069] In one example, a language model can be used to generate or correct representations, such as high-definition (HD) maps. HD maps are typically a type of map used for tasks such as autonomous driving, which may contain details or information that are not typically included in or associated with conventional maps. In an example HD map, individual segments of a road are individually encoded. For example, these encodings can distinguish areas in an intersection that are damaged to different lanes, and potential options for navigating on these lanes. Such information is helpful in intersections, where there may not be a map or clear lane markings for each available lane in each direction. This information helps the navigation system work more like a human would, with the ability to understand implicit information based on context, but with previous systems, these aspects need to be hard-coded, and thus limited in scope and difficult to expand. Each feature in the road can be represented by a node in a graph associated with an HD map. The language model can take this information and can make corrections or additions based on its understanding of the relationships and semantics of the environment, which can account for implicit reasoning that is typically performed by humans and that may otherwise be difficult to design or instruct automated processes to perform. While encoding aspects such as critical road boundaries can be relatively simple using manual approaches (at least for certain environments), encoding more implicit operations such as how to maneuver relative to crosswalks in complex urban environments (where options may vary based on the number and location of people at the crosswalk at any given time, the state of crossing signals that may not all be visible, and people's decisions to not follow the rules and cross the street at traffic lights or to jaywalk) or how to navigate through roundabouts or uniquely constructed areas can be much more difficult for traditional systems and can benefit from reasoning and similar determinations that can be performed using the language models presented in this article.

[0070] In at least one embodiment, a vector-based search can be used to find similar vectors (or other embeddings or encodings, etc.), such as those that may correspond to very similar intersections, for inference of various aspects of an environment or actions to be taken for such an environment, particularly where the raw sensor data may be inaccurate or incomplete. Therefore, this approach is not a rule-based search, but rather a similarity-based search based on what can be determined about an object or environment. Any appropriate vector-based similarity method can be used, such as a cosine similarity based on L1 similarity between feature vectors that can be attempted to be determined. For similarity-based searches, examples are provided that will be used for searching, rather than descriptions, labels, or queries that attempt to clarify aspects or parameters of the search. For similarity-based searches, an example is provided, and the search attempts to locate similar examples independently of any description, label, or annotation associated with the example (although in at least some embodiments, such information may be used as an additional search parameter). Such searches can be performed quickly using incomplete or slightly inaccurate information, which can provide significant advantages over rule-based searches, and can allow results that are not exact matches. Embedding can also represent various uncertainties about an object. Such embeddings can also be very lightweight representations that require little storage capacity or bandwidth, streaming, or transmission. As additional information about an object is obtained, the embedding of the object may also be updated.

[0071] In many cases, the language used with such models is somewhat lossy, so in at least one embodiment, it may be important to try to encode the features of such language in a way that retains as much important information as possible. For example, an image may contain a lot of details about a person or vehicle, and many of these details are lost in a compact language description. For many reasons, this is acceptable because information about a person's overall appearance or clothing does not typically affect the decision a vehicle makes about that person, such as avoiding entering within three feet of that person at any time. Therefore, the type of object can be used as the primary indicator or type of the tag, but additional details or information can be stored in the tag or tag descriptor in the tokenized text string. In at least one embodiment, geographic coordinates are also stored, so that in addition to information about connections or relationships, all nodes or tags have well-defined locations in space. Therefore, in addition to information about semantics and topology, nodes also store information about geometric structure. This information can also be general enough to support multiple domains or tasks that may involve similar types of objects.

[0072] In at least one embodiment, such additional information can be generated using conventional algorithms or machine learning, and other such options. For example, one or more machine learning models can be trained and used to provide information about the semantics, topology, and geometry (or other such aspects) of an object or environment. This can include the use of one or more language models that can take various types of inputs and output text descriptions or text content of any of these aspects. In some embodiments, raw sensor data can be provided as input, while in other embodiments, there can be at least a certain amount of preprocessing, such as to determine a bounding box around an object and extract relevant image data, or to perform basic object classification based on operations such as computer vision-based analysis, and other such options.

[0073] Figure 2BAn example tokenized text string 250 that can be generated for an environment according to at least one embodiment is shown. As shown, the string is a tokenized and continuous text string that represents individual objects as tags, similar to the nodes of a map graphic. In this example, there are various types of objects represented by tags and tag identifiers in the text string, such as signs, lanes, and vehicles. The text string also includes tag descriptors, or additional information for these tags. These descriptors provide additional information, such as the direction of traffic for a lane or information identifying connecting lanes. Other information is also encoded, which may involve key points or geometric structures of each tag about the environment. For example, key points (as indicated by letter pairs in a string for encoding geometric coordinates in the environment) can be used to indicate the geometric coordinates or boundaries of a lane or lane segment. Although the example text string is quite long (such as thousands of tags that can include a single environment), the text string provides the information necessary to perform tasks such as navigation or driver assistance in a much more compact form than the data is a set of high-resolution images or high-density point cloud representations of the environment. Although a single text string is shown, it should be understood that in at least one embodiment, multiple text strings can be generated to represent different parts or features of the environment. In addition, in certain embodiments, different types of information can be used with text strings, where appropriate for a given environment. In this example, the text string is generated using a specific language, such as RTL. The language is constructed so that the text string will be discreet and sequential in its tokenized (one-dimensional) representation. In at least one embodiment, the generated text string can be autoregressive, because the individual tags in the string will depend in part on one or more previous tags in the string. As previously described, unsupervised (or self-supervised) methods can be used to train language models so that a wide variety of concepts required can be covered without requiring a very large and varied corpus of annotated training data. In at least one embodiment, even if the text string is tokenized and sequential, few other structural restrictions can be imposed on the generation of the string to prevent these restrictions from becoming a bottleneck that may negatively affect performance due in part to the large amount of input data that may need to be processed. The structure may be flexible, similar to how there can be many effective ways to flatten all equally effective map graphs or object graphs. There may also be many different effective object graphs to represent the same environment, and the generation of tokenized text strings can have similar flexibility. This flexibility also helps text string generation to be better updated over time, as well as expanded to include a larger or smaller number of tokens based at least in part on changes in the relevant environment.

[0074] In at least one embodiment, a tokenized description generated using a trained language model can be used to provide an accurate representation of at least a portion of an environment. While such descriptions can be used for various "offline" purposes, such as simulation and environment reconstruction as described herein, the tokenized descriptions can also be used to make "online" or other (near) real-time decisions, such as when performing operations such as vehicle navigation (autonomous, semi-autonomous, or other) or robotic automation control or training. When performing or planning such operations, it may be important to obtain an accurate and robust understanding or representation of the environment in which the operations are to be performed. This understanding not only helps to ensure appropriate or expected performance, such as to accurately identify objects being interacted with and their precise locations, but also helps to avoid unexpected occurrences, such as those that may involve undesirable interactions with one or more objects in the environment.

[0075] For example, for operations related to vehicle navigation, the control system will typically include at least some type of perception system, module, or process. For example, the perception module may attempt to process observations or other such data about the environment (which may be received from or captured using one or more sensors in communication with the perception module) and attempt to "sense" information about the environment, such as to identify objects in the environment and determine information about those objects, such as their type, location, orientation, motion, etc. At least some systems will attempt to associate the perception data with map data of at least a nearby or adjacent portion or area of ​​the environment. Being able to associate map data with the perception data may help to improve the accuracy of the perception data by matching the perception data with "known" map information (such as the location of lane boundaries or intersections).

[0076] When one or more such maps are available, the perception module can utilize them by first determining the precise location information of the "self" vehicle being controlled, or at least being provided with navigation or operating instructions or information. In order to obtain precise location information, systems such as the Global Positioning System (GPS) and other such options can be used. However, there may be problems in obtaining accurate location information from such systems, such as when the GPS signal is unreliable or lacks sufficient signal strength. There may also be problems with the accuracy and / or current state of the map, which may also make it difficult to align the map data with the perception data. If the map data cannot be accurately aligned with the perception data, or if there is no sufficiently accurate map data available, the perception module may be responsible for detecting all driving-related objects around the self-vehicle in real time and with high accuracy, which may result in greater pressure on the perception module in terms of performance. This additional pressure may not only require and / or consume additional resources, but may also produce less accurate results or require additional delays, which may be undesirable for many intended operations.

[0077] Methods according to various embodiments may overcome at least some of these and other such deficiencies in existing solutions, at least in part, by providing enhancements to perception data using local map information. For example, such methods may provide robust positioning of an ego vehicle or machine without requiring high-end geolocation equipment and / or services, which may be expensive and may be unreliable based on factors such as location and signal strength. As an example, Figure 3 An example environment representation generation or reconstruction system 300 is shown that can be used in accordance with at least one embodiment. It should be understood that reference numbers can be transferred between the drawings of similar elements, but such use should not be interpreted as limiting the scope of the various embodiments. In this example, the capture device 202 includes sensors 204, 206, which use, for example, Figure 2A The described methods capture sensor data (or obtain other such observations) related to the environment 208, although other mechanisms or methods may be used within the scope of various embodiments to obtain such data. In addition, additional data may be used to attempt to sense information about at least a portion of the environment 208, as discussed in more detail elsewhere herein.

[0078] In this example, sensor data (and potentially other observations) from capture device 202 is provided to perception module 302. Capture device 202 may perform at least some amount of processing on the sensor data before providing it to perception module 302, which may include noise reduction, aggregation, correlation, redundant data point removal, etc. Sensor data may be provided in any suitable form, which may include image data, 3D point cloud data, feature vectors, etc. Another benefit is that this approach can enhance system robustness, including the ability to identify and resolve individual sensor inaccuracies or even failures. For example, perception module 302 may perform at least some amount of processing on data from individual sensors, and may determine when data from one sensor is unreliable, and then determine whether any unreliable data should be modified, weighted lower, or discarded. In at least one embodiment, perception module 302 may attempt to correct the values ​​received from individual sensors based on data from other sensors or related sources, as well as other such options.

[0079] The perception module 302 in at least one embodiment can perform tasks including those discussed in more detail elsewhere herein, such as extracting features from sensor data and attempting to identify objects in the environment, and determining relevant information about these objects. In at least one embodiment, the encoder can perform feature extraction or feature reasoning to extract features that may be relatively low-level and may not have additional explicit semantic meanings and encode them. These features can be used to generate relatively general and / or general (generic) representations of sensor data. Sensor data or perception data can be interpreted and / or associated in the cross-attention layer of one or more trained models. Such a model can try to associate relevant features to allow objects to be represented using shapes such as shapes that can be composed of lines, triangles, or polygons, and semantic information can be identified and associated with the represented objects. In at least one embodiment, the model can analyze feature vectors including appearance information of an object without any higher-level structural information, and try to determine various properties related to semantics, relationships, topology, geometry, etc. Therefore, the encoder can only try to use a subset of points or embeddings to represent sensor data as faithfully and accurately as possible in a way that is friendly to downstream processing. A model (or other such component) that receives these features or embeddings can then attempt to understand these encodings using domain-specific knowledge.

[0080] The sensor data can be extracted and / or encoded in a variety of different ways. For example, there can be an encoder for each sensor so that each sensor can output a corresponding label stream that can be input into the model. The trained model can then fuse the information in the parallel streams with the map data discussed herein. In other embodiments, sensor data fusion can be performed before generating the label stream. For example, data captured by sensors around the vehicle can be used to generate a point cloud representation of the vehicle's surroundings, and this point cloud representation can be analyzed to generate a label stream. The association of sensor data can be performed in such a way that in many cases the calibration information of the sensor may already be available so that the position data can be determined relative to a consistent coordinate system or reference system. Then, in at least one embodiment, the model can analyze the consistent 3D representation to generate a single representative label stream. The association of sensor data can also solve problems related to multimodality, because any of many different methods can be used to interpret and associate data from different types of sensors. For example, algorithms that can associate appearance features extracted from camera images with position data obtained from a lidar system or the like are available.

[0081] In at least one embodiment, the perception module 302 may attempt to identify only specific objects or object types of interest in order to reduce environmental perception to a more manageable task. For example, for navigation of a vehicle, this may include detecting static and / or dynamic objects related to driving, such as lane boundaries, traffic signals, other vehicles on the road, pedestrians within the vehicle range, etc. The perception module may determine that there are static objects away from the road, or objects that are unlikely to affect navigation, and may classify these objects as unimportant objects or exclude these objects from recognition, as well as other such options. In at least one embodiment, if it is not an absolute position relative to a certain geographic origin or reference plane, point or coordinate system, the perception module 302 will attempt to determine at least the relevant position of a specific type of object relative to the ego vehicle. For objects that may be in motion, such as vehicles on the same road as the ego vehicle, this may include a position at a specific point in time or a range of positions within a time window, such as a window with a length corresponding to the capture rate or refresh rate of the relevant sensor used to determine the position. For objects in motion, the perception module 302 may also attempt to determine the direction and / or rate of motion, such as speed, acceleration or deceleration, which may be based in part on the position or motion information determined for a previous point in time or time window. In at least one embodiment, the perception module 302 can generate an accurate reconstruction or representation of the environment in which the ego vehicle is operating so as to allow the vehicle control system 316, process or module to determine instructions for safely operating the vehicle within the environment to achieve desired goals, such as to safely navigate the vehicle to a target destination.

[0082] As previously described, in at least some embodiments, vehicles such as autonomous or semi-autonomous vehicles can operate based on such perception data. However, in order to provide a more accurate perception of at least relevant portions of the environment 208, a system according to at least one embodiment can attempt to use local map information to enhance or improve the accuracy of such perception data. Figure 3In the example environment representation generation or reconstruction system 300, the map module 308 can access map data stored in a map repository 310 or other such location. The map repository 310 can be available on the vehicle or accessible via a wireless data connection, for example, where relevant map data can be pre-acquired by the vehicle based on the current and / or expected location of the vehicle, such as a given minimum distance for the vehicle or along a current navigation route. Pre-acquisition can be used to try to ensure that relevant map data is available even if the wireless network connection is weak, unstable or otherwise unreliable or unavailable in a given location or area. The map module 308 in this example can work with the positioning module 304 to try to determine the current geographic location of the vehicle. The positioning module 304 can include at least one system, sensor, device, component, process, service or other such mechanism, or communicate with at least one system, sensor, device, component, process, service or other such mechanism to determine the location of the self-vehicle. This can include, for example, using a GPS (or GNSS) system 306 that uses satellite-based radio signals to perform geolocation anywhere a sufficiently strong signal can be received from at least a minimum number of satellites (such as at least three or four satellites). One advantage of GPS is that it can be highly accurate, does not require outgoing data transmission from the ego vehicle, and does not require an active network connection, such as the Internet or a cellular connection. GPS systems generally must have an unobstructed transmission path with a minimum number of satellites, however, this may not be possible in some places, such as cities with tall buildings, tunnels, or mountainous areas. Other geolocation mechanisms may also be used in other embodiments, such as those that are determined at least in part based on signals sent from the earth or identifiable features in the nearby environment 208, as well as other such options. GPS receivers are typically on the vehicle, while other methods may use components that are not on the vehicle, but latency and connectivity may become a problem in some cases. In at least one embodiment, the positioning module 304 may attempt to use other available information to improve or stabilize the location data from the GPS (or other such system), such as the vehicle's speed and direction of travel, the location of nearby objects, signal noise reduction, etc. In this example, the map module 308 can receive geolocation data from the positioning module 304 and can determine the current location of the ego vehicle relative to the stored map data. In at least one embodiment, this can be used to obtain and / or pre-acquire local map data for the current geographic location of the ego vehicle.

[0083] At least a selected portion of the perception data from the perception module 302, and the geographic location and / or local map data from the map module 308, can be provided as input to the alignment module 312. In at least one embodiment, the alignment module 312 (and similarly, the positioning module as described above) can be a dedicated or independent component or process, or can be part of a system component that also supports other functions, among other such options. Such an alignment module 312 can attempt to "align" the perception data with the map data to provide a more accurate and reliable interpretation of the location of the ego vehicle and the surrounding environment. Alignment methods can include various SLAM-based methods, for example, in which data obtained from one or more perception modules and / or sensors (such as cameras or lidar sensors) are used to perform simultaneous localization and mapping. Such methods can be used to attempt to detect, classify, and track objects in a dynamic environment based in part on the movement of the ego vehicle and / or detected objects. However, as previously described, such methods generally require highly accurate geolocation and / or map data and sufficient perception data to allow proper alignment. Issues such as obstructive objects or features in the environment can also prevent reporting of perception data with a sufficient level of confidence.

[0084] The method according to at least one embodiment allows the alignment of available map data with perception data, which can provide more robust positioning than various existing methods. In at least one embodiment, local map information is retrieved to find an approximate geographic location, such as can be determined using a positioning module. Then, the local map data can be matched and aligned with the perception results based in part on, for example, static landmark features (such as traffic signs, traffic signals, poles, crosswalks, stop and yield lines, lane dividers or curbs, and other such options). Such alignment can lead to robust positioning of the self-vehicle, such as in the case where GPS signals or data may be unreliable or unavailable. This method avoids the need for a highly accurate location determination component because the general location is sufficient to associate or align the local map data with the content identified in the perception data. For example, the location in an urban environment can be determined only within a four-block area, but the static objects identified in the perception data can be used to more accurately determine the location of the vehicle within the four-block area. Once the location is correctly determined and the local map data is aligned with the perception data, those data can be considered together to generate a more accurate and consistent representation of at least this part of the environment. In some systems, position determination may not be available at all, but past geolocation information and motion information about the vehicle may be used to infer an approximate position, and localized map data for that approximate position may then be aligned with the current or most recent perception data.

[0085] In at least one embodiment, the aligned map data can be used as prior information of the surrounding environment. The obtained sensor data or other observations can be fused together and used to detect or reason about static and / or dynamic objects. These results can be regarded as likelihood functions of these objects given the surrounding environment. A language model or other language-based generative model (such as an LLM) can take the aligned map data and the detection result data as input and can produce an updated description of the surrounding environment. In at least one embodiment, these descriptions can be in the format of a tokenized description of a domain-specific language (such as RTL) or a text-based tagged string. The geographic location of the self-vehicle can also be updated based in part on the updated information.

[0086] Such a framework can provide several advantages over previous solutions. For example, such a framework can organically fuse map and perception modules with neural networks (such as, for example, large) language models, which can reinforce the spatial consistency and semantic structure learned from map data. Map information can help the perception module reason about occluded map features and enhance ambiguous map features that the perception module's detection algorithms or processes may not reliably report. One advantage of using language models is that these models can be trained using large amounts of map data, which includes not only the final, accurate map, but also intermediate data about how human operators correct errors in initial maps. Language models can use domain-specific language to learn the natural features of natural maps, and this trained language model can then be used to help perform perception-related tasks.

[0087] In at least one embodiment, for example, the language model 314 can be trained to understand aspects such as road entities and relationships, and can use this understanding to reason and / or correct relationships, as well as reason about where entities should exist. This knowledge can be built in part by submitting millions of examples of valid map documents to the deep learning model, which contain road representations in a tokenized representation of a domain-specific language (such as RTL). In many cases, a given map document will only encode a small, localized portion of the overall map of the environment, where that portion may be related to intersections or other such features. This approach can be targeted at what a driver can conceptualize at a single moment in time based on what is visible to them. The language model can be trained to generate and / or complete these documents, such as providing at least a portion of the document as input, and the model can use its knowledge to fill in any gaps, make appropriate corrections, or otherwise enhance the document.

[0088] After at least initial training using this approach, the language model 314 can understand domain-specific language and what constitutes a valid and / or complete map. Such a model can then be applied to a variety of domain-related tasks, which can include improving perception data of the environment. Perception in general can be a very complex task, but its use for tasks such as autonomous navigation can lead to additional complexity, where perception must be highly accurate in real time. Figure 3 As shown, the perception module 302 can attempt to understand or map its environment based at least in part on data from a set of sensor inputs, which can include a single camera frame, a video clip from a single camera, or up to a sequence of frames from several sensors of various types (e.g., cameras, lidar, radar, or sonar), among other such options. Perception modules are typically trained by providing example inputs and outputs, but training perception models in this manner becomes increasingly difficult as the complexity of the outputs increases. In addition, complete or usable maps such as high-definition maps are typically quite complex and difficult to generate accurately and quickly.

[0089] The method according to at least one embodiment can generate or enhance such a map in (near) real time, in part by representing the map data using tokenized descriptions, which are lightweight and can be processed using a trained language model. Such a model can generate or enhance tokenized descriptions that are significantly more complex than the descriptions generated by typical perception models that focus on limited tasks (such as identifying lane lines or traffic signs in a portion of the environment). The language model according to at least one embodiment can perform these and other such tasks, including organizing the perceived objects relationally and topologically and annotating them with rules as representations of roads, such as autonomous vehicle systems can operate on them. For autonomous vehicles to operate, in many cases, the vehicle will need to understand the concept of a road (or other navigation area), including the rules and current state of the road. The use of a separate language model 314 can help greatly improve the quality of the resulting perception by leveraging an understanding of one or more relevant maps of the environment, including the road to be navigated. In at least one embodiment, the sensor deep learning backbone can focus on sensor feature reasoning, and one or more sensor-to-map layers can focus on how to organize these features into a map. The language model 314 can then advance toward a map that is consistent with how the map is understood to work, and can guide subsequent sensor-to-map layers toward a consistent and coherent map representation of the observed road. This approach provides advantages over previous approaches that rely heavily on existing HD maps to perform L4 navigation tasks, for example, using sensor data primarily to detect and avoid dynamic objects such as pedestrians or other vehicles. Other approaches attempt to perform a mapless (or primarily mapless) type of perception, in which both static and dynamic obstacles are detected from sensor data, but the accuracy of these approaches may be limited due to factors such as line of sight limitations and imperfect environmental conditions, and limitations on accurately determining complex information in real time due in part to limited processing capacity, memory, and other such factors. The methods according to the various embodiments presented herein can utilize map data (if available) to provide more accurate and robust perception determinations. In various embodiments, map data at various levels of detail and / or from various sources or providers can also be aggregated and utilized.

[0090] like Figure 3As shown, the trained language model 314 can take both the perception data and the aligned map data as input. As discussed in more detail elsewhere herein, other types of inputs may also be provided. Aligned map data can help improve the accuracy of the perception data, as well as potentially enhance the perception data with information that may not be perceived from the sensor data alone, or enhance the perception data with information that may not be perceived with sufficient confidence without confirmation from the aligned map. The ability to use a lightweight, compact, and robust representation of perception and map data (such as a representation in the form of a tokenized description of a domain-specific language) can allow for more complete and complex perception results of the environment to be generated in real time as needed for autonomous operation when computing resources (such as onboard an autonomous vehicle) may be limited. This approach can also allow for natural integration with current perception stacks, which in at least one embodiment can be based on converters. The ability to use a language model to enhance perception data can also allow for a more conceptual understanding of sensor data and information inferred or perceived from the sensor data. Current perception modules do not provision or provide a higher-level understanding of the relationships or semantics of individual objects in the environment and the relationships or semantics between them, and are therefore limited in their ability to provide complete and robust perception determination.

[0091] In addition, the ability to leverage existing map data in the perception system helps reduce the processing capacity and time otherwise required to generate perception determinations based solely on sensor data. For example, if a reliable map (which can be quickly associated with features in the perception data) is available, information related to lane and intersection locations and relationships can be used without having to use time and resources to try to infer these locations and relationships from sensor data. Such use can help reduce the burden on the perception stack. In addition, relationship and semantic information can be more easily provided in map data generated offline rather than in perception data that needs to be generated in real time using limited resources, so the ability to leverage map data can allow such information to be effectively provided as input to the language model. In at least one embodiment, map data can be viewed as a very dense representation of at least a portion of the surrounding environment. It can then be assumed that many fine details about the environment will not contradict the sensor data, at least to some extent. In addition to providing at least some of the perception data without the need for detailed processing of the sensor data, such information can also help enhance the perception data with details or information that may not be determined from the sensor data. Using a tokenized description to store such information also helps retain such additional information because the description is lightweight and compact relative to other environmental representations, and can be sparse and more discrete. However, in at least one embodiment, the information to be stored in such a representation can be limited to information that is useful or even necessary for the task to be performed (such as autonomous navigation). By excluding information that is not required for such a task, the tokenized description can be more compact and allow faster processing. As previously described, the tokenized description can correspond to the type of object graph, where each object or node can correspond to multiple tags indicating information about the object or node. In order to keep the tokenized description as compact as possible, the language model can be trained to generate tags only for certain types of objects, and to generate certain types of tags only for specific types of objects. For example, a language model can generate a representation that includes information about objects such as roads but does not include information about objects such as sidewalks or courtyards. The model can also generate tags for only specific aspects of the road, such as information about lane dividers or road boundaries, but not information about speed bumps or cattle fences in the road. This approach allows a large amount of data to be selected (making the process more robust and accurate than existing methods), and also allows only the portions of this data that are relevant to a specific task or operation to be selected in order to reduce the processing, storage and / or transmission resources required for the tokenized representation. Storing map and perception data in a robust yet compact representation also allows for faster and more direct fusion of data in limited resource capacity environments and / or in real time.The ability to perceive and associate objects using language-based rules from trained models rather than hard-coded heuristics as in previous systems also helps avoid problems where perception is brittle or easily stuck in local minima or other situations where acceptable determinations cannot be identified. Augmenting perception data with map data and model learning helps to properly interpret, correct, and enhance perception data to allow accurate determinations to be made in situations where sensor data may lack the data required for correct or acceptable perception. For example, in situations where snow covers the ground and features such as lane markings are not captured or represented in the sensor data, the ability to associate map data with detectable features in the environment and model learning based on similar situations can allow the vehicle to operate even when the sensor data alone is insufficient to allow continued operation in a rule-based or heuristic-based system.

[0092] The language model can also be flexible in the way it aggregates information from the input map and / or perception data, because the language model can use a larger context window. With more information available, the language model can select important information based on relationships, semantics, or other such aspects, without being constrained by specific rules or fixed values, such as only considering objects within 30 feet of the vehicle, or objects that will be in a specified vicinity of the vehicle within the next 30 seconds, etc. The ability to obtain and use additional information as needed can also make tokenized descriptions more expressive and robust than other perception representations.

[0093] In at least one embodiment, the map data may be represented by a tokenized description that stores information about objects in a nearby portion of the environment that are relevant to a specific task. For vehicle navigation, these may include objects related to driving. The information may be encoded using a domain-specific language such as RTL. The perception data generated by the perception module 302 may be in different formats, such as a very dense but low-level representation of a portion of the environment in a 2D bird's-eye view. The perception data may not include any semantic labels or at least a full set of semantic labels for objects in the environment, while the map data may already include relevant semantic labels. In order to be able to apply labels from the map data to the perception data, the map data needs to be aligned with the perception data. Both the aligned map data and the perception data may be provided as input to the language model 314, which may perform association, apply semantics, relationships, and other relevant information or aspects, and generate a more detailed tokenized description of the environment, which may be the same or a different domain-specific language. In at least one embodiment, the matching or association of the map data with the perception data may be performed implicitly by the cross-attention layer of the language model 314. The output enhanced perception data may then be a tokenized description of the environment that has been generated based in part on the fusion of map data with real-time perception data. This may be thought of as prompting the language model to describe the sensor input in a manner consistent with the map data. The enhanced tokenized description may then be a more complete and accurate representation of the surrounding environment than the map or perception data alone, while still being a relatively lightweight discrete and compact representation.

[0094] In at least one embodiment, the tokenized description or representation does not need to include decoded information in each tag. For example, an object may be associated with many tags that store information of different aspects of the tags. In some embodiments, the text within a given tag of a tag string or sequence can be a human-readable language and easy to be understood by a human reader. In some embodiments, the information in each tag can also or alternatively be in the form of a text representation, but not in the form of a human-readable language. For example, text encoding can be used for these various aspects, and the text encoding can be stored in each tag. A human reader accessing the tag may not be able to interpret the text information, but such information can be understood by a system that can decode and / or understand the text encoding and a language model for generating and / or processing the text encoding of various tags. In some embodiments, for example, the final tokenized representation can be a domain-specific language that a human reader should understand, but an intermediate or other such tokenized description or representation can be in other languages, such as can include encoding or other such text components. In at least one embodiment, the model can be trained to use encoded tags in a domain-specific language that is understandable to humans as input and output tags in a domain-specific language, as well as other such options. As mentioned above, the language model may also take various other forms of input, such as including features and / or embedding vectors, points in a latent space, raw sensor data, etc. The final output representation in the domain-specific language may then be used to perform various operations, such as providing a visualization of at least a portion of an environment in a 3D space.

[0095] In some cases, there may be conflicts between map data and perception data. For example, an obstacle may prevent sensor data from capturing an object represented in the map data. Similarly, there may be objects detected by sensors that are not represented in the map data. In addition, there may have been changes or updates to the environment since the map data was generated or updated, such as may be related to temporary changes (such as may be due to construction or events) or more permanent changes (such as adding another lane on the road or modifying the turning lane at an intersection), as well as other such options. One advantage of using a trained model (such as a language model) to align and fuse map and perception data is that a large set of explicit rules is not required to handle the wide variety of potential conflicts that may arise. The language model can be trained to output the results that the model considers to be the most consistent and / or most likely based at least in part on the received input data. This approach is particularly beneficial for operations such as automatic change detection, because the trained language model can identify any changes and then generate an updated map (or updated map dataset) representing the identified changes with high accuracy and high probability. This approach can help ensure the reliability and up-to-dateness of map data without any manual updates, reviews, or interventions. In the event of a conflict (which may or may not be due to an update), the trained model can generate the most likely and consistent version based on the available data and the model's own learning.

[0096] In at least one embodiment, a model for aligning map data with sensor data can perform a type of implicit localization in which the model can attempt to use its learning to infer the most likely alignment and / or localization result. A language model that receives input in the form of a tokenized description as described herein can benefit from including semantics, relations, topology, geometry, and other such information in the representation to better infer consistent outputs. For example, if the map data and sensor data represent a turning lane that starts at different locations, the language model can process the map data and sensor data using relational information about other objects in the scene (such as the locations of signs, signals, and other vehicles) to infer the most likely starting point of the lane based in part on its learning from other similar roads and environments. While sensor data may be more up to date than map data, sensor data is also more susceptible to factors such as noise, environmental conditions, obstacles, and the like. The model can take all available information, including semantics and relations, and infer the most likely representation based on the information and its learning from training data at similar locations.

[0097] In many cases, there will be dynamic objects represented in the sensor data or perception data that are not represented in the map data. This can include various types of objects, such as vehicles, pedestrians, scooters, etc. The model can still attempt to make inferences that are consistent with the map data, even for objects that are only represented in the perception data, which can be based in part on the various types of objects that are generally expected to be located in specific locations, such as the operation of intersections along roads or crosswalks. In addition to basic information about the object (such as type and location), the model can also attempt to encode information about the dynamic properties of the object, such as speed, orientation, direction, acceleration, deceleration, or other such information. In at least one embodiment, these various aspects can correspond to encodings in one or more additional tags associated with a given object. In at least one embodiment, the model can maintain a consistent reference frame for determining dynamic data so that the data is consistently represented in various tags over time. Storing dynamic information can increase the number of tags required, such as 2000 or 3000 tags for an intersection instead of 1000, but the lightweight nature of the tokenized description allows such additional data, and the description can be kept discrete and compact by storing only that portion of the dynamic information that is relevant to the operation to be performed and only for those objects determined to be relevant to the operation over an upcoming time period, distance, or other such factor. In at least one embodiment, an adaptive representation of spatial location can be used to reduce the amount of precision required and allow fewer and / or smaller tags. Other optimizations can also be performed, such as to output tags with critical information earlier in the string than less critical information, which allows faster processing of that information.

[0098] For example, to determine the motion aspects of a dynamic object, the trained model may have the ability to analyze information across multiple sensor data frames or time points. For example, this may include a cross-attention neural network layer analyzing data across multiple timestamps, such as the most recent timestamp and multiple previous timestamps across a context window of any appropriate length, which may be limited only by the amount of available data and the capacity of system resources. The model may attempt to extract or identify information from any or all of these timestamps that is determined to be relevant to determining current information related to a dynamic object. For example, data from the current and immediately preceding timestamps may be required to determine the current speed and direction of motion of an object, while data from at least 3 timestamps may be required to determine the motion orientation or motion path. The model may determine what information and how much information to obtain from the current frame and any identified previous frames in order to determine the dynamic information of the current timestamp. Although using data from a larger number of timestamps can help provide a more accurate determination of aspects such as motion paths or motion styles, data from previous frames that are further away in time will generally be less useful when determining information such as current speed or acceleration, which may change over time, making older information less relevant. The trained model can then aggregate the corresponding dynamic object data based on key values ​​such as attention. Thus, the model is not only able to identify the data required to calculate the value of each dynamic data across multiple timestamps, but also able to determine which of the available data points is most relevant to calculate the accurate value. In at least one embodiment, the model is not trained to fuse this information in a specific way, but can learn from the training data how to correctly fuse specific types of information together to generate accurate and useful results.

[0099] At least relative to various other systems, the ability to extract and fuse data over large context windows can be improved, in part because the information processing is performed in a latent space. Such an approach can be more efficient and lightweight, without requiring any sensor input to be discarded, but rather allowing the model to be selective about which subset of the data the model determines to be relevant. For example, a model may extract information that identifies the location and orientation of trash cans along the side of a road, but may not care about information such as the color, type, or design of the trash cans. The relevant information can be extracted, fused, and used to generate appropriately labeled information to be written into a tokenized description, which will be generated for the current state of the environment (or other relevant portion of the environment) in which the ego vehicle is located. The resulting description is both conceptually and spatially compact because the description only contains information relevant to a specific task or operation, such as vehicle navigation or operation in the environment. The description can further include information relevant only to the specific task or operation at the current point in time, so that, for example, information about vehicles within a radius of approximately 500 feet can be considered, but information about vehicles beyond such a distance can be excluded, where the distance can depend in part on factors such as the current travel rate or location type. This optimization can be achieved through filtering, model training, or a combination thereof.

[0100] In at least one embodiment, a large language model can be used to fuse the perception data with the fusion data. The language model can be trained using domain-specific data to generate a tokenized description in a domain-specific language. The language model can be trained to identify and extract relevant information from both the map data and the fusion data, and fuse the information together in a manner that provides the most consistent representation. This method can be used in a variety of fields, such as for driving, navigation, robotics, or automation. The method can also be used to generate tokenized descriptions for various purposes, such as operations, simulations, or training, where the tags included in the tokenized descriptions generated by the trained model may vary for different purposes because the important or relevant aspects of various objects may also vary. Therefore, this method can allow simplicity of architecture and scalability of data, as well as other such advantages. The model trained as described herein can be compact and lightweight enough to allow online processing of real-time operations, and can provide tokenized descriptions that can be used offline by downstream models, such as for prediction, planning, training, and other such operations.

[0101] In some cases, sensor data may not be sufficient for expected operations, such as vehicle navigation and operation. For example, there may be obstacles that prevent data from being captured, noise in the data, problems with sensor calibration or availability, and various other such factors. In at least some systems, the lack of reliable sensor data will hinder successful operation execution. The method according to various embodiments can overcome at least some of these problems of sensor or perception data quality by trying to infer the most reasonable and consistent representation of the environment based on at least available map data and any previous data of the environment and any sensor or perception data received. When training such a model, there may be a situation where a portion of the perception or sensor data is removed before training, or only the map data is provided. In some cases, data from one or more sensors can be excluded from consideration during training. In other cases, all perception data may be available, but at least a portion (if not all) of the map data can be excluded from consideration. It may also be the case that only lower resolution maps such as SD maps are available, while higher resolution maps such as HD maps are not available. This method can train a language model to infer the most reasonable and consistent representation based on the data available at any given time. This approach can help avoid situations where an action cannot be performed due to issues with available data, such as a situation where an autonomous vehicle is stopped in the middle of an intersection because the available data is insufficient for the vehicle control system to make any action decisions that meet a minimum confidence or safety threshold.

[0102] In addition to real-time operations (such as autonomous vehicle navigation) that can be performed based on a combination of static and dynamic data, it is also possible to use such data to perform reasoning that can allow active or predictive operations to be implemented. For example, the model may be able to determine aspects such as the direction, speed, and acceleration of nearby vehicles. Such information, together with available map data for at least a portion of the environment, can allow the model to predict the state of the environment at a series of future timestamps even if sensor data has not yet been received. Being able to predict the future state of the environment (including dynamic objects) allows prediction of undesirable actions or states, such as other operating states that may be related to collisions or may be below a minimum safety threshold or otherwise fail to meet at least one requirement for safe operation. The availability and understanding of semantic information and relationships can also help trained models make more accurate predictions or reasoning than might be possible using other methods or components. Instead of simply sending a tokenized description of the environment to a downstream component to attempt to perform operations such as collision avoidance, the model can reason or predict potential conflicts and send warnings or other indications of potential undesirable actions. Such warnings can be sent in the same tokenized description, or as a separate tokenized description of higher importance that can be sent to a different system, or directly to a specific system. In at least one embodiment, the model may also generate one or more recommended actions or instructions to try to avoid the predicted collision or other undesirable state. The ability to save even a little bit of latency by making predictions in the model itself rather than in downstream systems or processes can help improve the overall safety of operations, especially for high-risk operations such as high-speed autonomous navigation. This approach also has the advantages of model-based observation or learning, based on semantics and other information reasoned for objects in the environment, rather than rule-based approaches that require hard coding of various existing collision avoidance systems. Even under different environmental conditions, model-based approaches can be more flexible, more scalable, and easier and less complex to train than a set of rules that need to be highly accurate and comprehensive to encode, manage, and update. This approach also helps make better decisions for general operations, such as better determining when to change lanes to merge into or drive into traffic based on the predicted future state of the environment rather than currently available sensor data. This approach can help better perform operations for abnormal situations, such as when construction or accidents occur and someone is directing traffic. Unless you have specific rules for a given situation, it is difficult to correctly interpret all situations and make accurate decisions in a rule-based system, but a language model-based approach can determine the operating patterns of other vehicles and can infer one or more appropriate actions to take based on these patterns and the available environmental data, even if the vehicle control system has never encountered this situation before.

[0103] Figure 4AAn example process 400 is shown that can be performed to generate a labeled description of an area based in part on a fused map and perception data according to at least one embodiment. It should be understood that for this process and other processes described herein, unless otherwise specifically stated, additional, fewer or alternative steps or similar or alternative sequences, or at least partially in parallel, may be performed within the scope of various embodiments. In addition, although this example and other examples will be discussed herein with respect to driving or navigation domains and environments, other types of domains, environments and representations may also be used and / or generated within the scope of various embodiments. In this example process, a set of observations corresponding to a region (or a domain, region, physical location, or other part of a physical environment) is received 402 or otherwise obtained. The set of observations may include, for example, data captured by one or more sensors, or other observations or data representing objects or features in the environment. A set of representative features may be extracted from the set of observations and analyzed 404 to generate a set of perception data for the area. The perception data may include information about objects in the area that are identified or inferred based in part on the set of observations. In at least one embodiment, the set of observations may be obtained over a series of time points or timestamps and perception data may be generated or updated for each time point, where the perception data may vary based on the movement of one or more objects in the environment or the movement of the ego vehicle corresponding to the sensor used to capture the set of observations, and other such options.

[0104] In this example, the geographic location of the ego vehicle in the area may also be determined 406, and this determination may be performed in parallel with the generation of the perception data, such as to allow the geographic location data of each timestamp to be updated. The geographic location data may be determined using a location determination sensor, system, or mechanism (such as may be based on a GPS signal), and may be used with a positioning system to attempt to improve the accuracy of the location determination, such as may be based in part on the movement of the ego vehicle, sensor data about the environment, or other such information. Based at least in part on the geographic location, a set of local map data including the geographic location may be identified 408. The local map data may be selected from a larger map data set and may have a fixed size, or may have a size determined based on factors such as resource capacity, vehicle movement, map density, or other such factors. The local map data may then be provided to an alignment module, component, or process together with the perception data, which may associate the local map data with the perception data 410 to generate a set of aligned map data. As discussed, the alignment may help ensure that objects in the environment represented in both the map and the perception data are correctly identified and located.

[0105] In this example, the aligned map data can be provided 412 to the trained language model as input along with the perception data. The trained language model can analyze the aligned map data and the perception data to try to infer a consistent representation of the region, which includes information about the semantics, relationships, topology, and geometry of the region. A text-based representation of the region can be generated, such as a tokenized description that can be associated with one or more tags of the identified objects in the region, where a given object can have multiple associated tags that store text information about one or more aspects of the object, such as static or dynamic information that can be related to the object. A tokenized description of at least a portion of the region can then be received 414 as an output from the trained language model. As previously described, the tokenized description can benefit from other learning of language rules and language models. The tokenized description in this example can be provided 416 to the control system or other downstream modules or processes of the self-vehicle for determining one or more operations to be performed. For example, for an autonomous vehicle, this can include the operation of the vehicle in an upcoming time period, where the correct and / or safe operation of the vehicle is determined in part based on the tokenized description. The tokenized descriptions will be compact and robust, and full of important information, allowing highly accurate operational decisions to be made in real time.

[0106] The processes discussed herein may also be used for other purposes discussed and suggested herein. For example, this may include robotic or automated operation in an environment, as well as the re-creation or simulation of a physical environment, or the training of a control system to operate in such an environment. Figure 4BAn example process 450 is shown that can be performed according to at least one embodiment. In this example, a set of observations of a physical environment is obtained 452. These observations may include data captured by one or more sensors or systems at a current or past point in time. The set of observations may also or alternatively include sensory data determined in part based on the captured sensor data and / or other such information. The sensor may be located on an object to be operated in the environment, on other objects in the environment, or otherwise positioned relative to the environment. A set of map data of the physical environment aligned with the observations may be determined 454. This may include determining at least one reference location in the environment corresponding to the set, and selecting a portion of a larger set of map data including and / or surrounding the reference location. Local map data may be aligned with the set of observations so that the location and type of the object are associated and consistent between the map data and the observation data. Then, the aligned map data and at least a subset of the set of observations may be provided 456 as input to a trained language model. The trained language model may attempt to fuse the map data with the observation data in a consistent manner using semantics, relationships, topology, geometry, and other such data to infer a single representation of at least an area in the physical environment. The model can generate a text-based representation of the area, which can include a tokenized description that includes a sequence of text tags corresponding to objects in the environment. The tokenized description can then be received 458 as an output of the trained language model. The tokenized description can then be provided to downstream components, which can perform online or offline processing to operate in real time if necessary. As previously described, such a tokenized description can be used for a variety of purposes and can be generated in at least one domain-specific language suitable for these purposes.

[0107] It should be understood that representations, such as feature vectors or tokenized descriptions, of areas of different types or sizes may be generated. For example, a tokenized description may represent a single, larger intersection, and / or there may be several other tokenized descriptions used to represent portions of an intersection, such as individual lanes or traffic signals associated with the intersection. In at least some embodiments, the level and type may be user-configurable, but in at least some embodiments, tokenized descriptions should be generated for a set of domains where specific rules or behaviors exist or where having more detailed data may be important or beneficial. For example, a long two-lane highway in the desert may benefit from few, if any, tokenized representations, as a vehicle may primarily view the location as a long lane in a given direction and navigate primarily using sensor data. For example, for very complex intersections or roundabouts, it may be beneficial to have one tokenized description of the large roundabout and associated entrances and exits as a whole, but it may be beneficial to have a more fine-grained representation of portions of the intersection, where individual entrances and exits may not be highly similar, may have different expected behaviors, etc.

[0108] In some embodiments, the vector database or latent space management algorithm may attempt to remove redundancy, such as points or vectors from separate sources but referring to the same area or domain. In some embodiments, highly similar domains may not have individual vectors or points stored, but may have a single representation stored associated with these various domains. For example, in a planned community, all intersections may be designed to be identical, but there will typically be at least some small variations. In at least one embodiment, a single feature vector associated with all of these intersections may be stored, rather than storing separate but highly similar feature vectors for each intersection.

[0109] In another example, a long highway may have many substantially similar sub-regions, at least from the perspective of features of interest. In this case, a single feature vector (or a small number of feature vectors) may be used to represent potentially many sub-regions along a very long stretch of highway, or a section of the highway may be used as a representative and then a feature vector representing a portion of the highway encoded. Instead of having hundreds or thousands of feature vectors that are highly similar or contain a large amount of shared information, a single representative feature vector may be used for the highway, or at least a large portion of the highway that is substantially similar, at least from the perspective of features of interest, and the single representative feature vector is then associated with multiple sub-regions or locations (or ranges of GPS coordinates, etc.). Different feature vectors may be encoded for specific off-road or other sections, but in many cases, each highway or road with a long, similar stretch may be represented by a single feature vector that can be reused appropriately.

[0110] By storing representative feature vectors for similar locations rather than feature vectors for each location, the size of the latent space or vector database can be significantly reduced, enabling large map representations of the entire area or environment to be stored on a vehicle or computing device rather than managing subsets associated with a specific time or location. In some embodiments, cluster density or proximity can also be used as a determining factor for reducing the number of representative points or vectors for similar areas. The amount of information stored in each tag of the tokenized description can also be variable or user-configurable, where some information is required for certain tasks or operations and additional information is optional. For example, if necessary, a description of objects near a road (e.g., buildings, postal boxes, etc.) can be stored, but the description may not be required for navigation or at least semi-autonomous navigation. In at least some embodiments, dimensionality reduction can be performed on the feature vectors in an attempt to further reduce the size and complexity of the vector database (or latent space, etc.). Such an approach can also help improve the efficiency of clustering and can result in fewer but more relevant clusters in at least some cases.

[0111] A method according to at least one embodiment can formulate a wide range of geospatial information processing and related tasks (such as map building, map editing, map-based navigation, planning and driving) as document manipulation tasks that utilize one or more LLMs to solve them in a unified and joint manner. In at least one embodiment, the LLMs are constructed using one or more deep neural networks (DNNs) that are trained using textual information (such as textual information in a domain-specific language (DSL) like RTL) and textual representations of geospatial information (such as maps). The DSL described herein may be referred to as, but not limited to, RTL. RTL may rely on the existence of a rich database of features and a graph describing the relationships of these or other such features within a map or other such representation. Using one or more automated processes or operations, a set of map features / landmarks (e.g., those encoded in an HD map using a data format suitable for HD maps) can be deterministically converted to RTL, and vice versa. In some embodiments, the graph may be represented as a knowledge graph expressing road objects, road object relationships, and road network topology, rather than general knowledge.

[0112] In some embodiments, the LLM (or other language model type) can retrieve and / or access map data or other information determined to be necessary to generate output using one or more application programming interfaces (APIs) and / or plug-ins (e.g., third-party plug-ins). For example, to retrieve additional context information, additional map information, additional feature information, and / or other information not directly included in the prompt to the model, in some embodiments, the system can use the LLM to generate one or more prompts or queries to one or more data sources (e.g., Open Street Map (OSM), wolfram alpha, local map databases, etc.) via one or more APIs or plug-ins to obtain the additional information required (or deemed necessary) to respond to the initial query or prompt. In at least some embodiments, this method of querying additional resources can be recursive, in that the system can continue to access one or more data sources via APIs and / or plug-ins until it is determined that the necessary information has been obtained, or until no additional information is available. In some embodiments, the initial prompt for the model can be generated using one or more APIs or plug-ins (such as an API for retrieving an RTL description of a selected portion of a map). For example, a user may select, via the API, a portion of a map to be processed or analyzed using the LLM, and the API may return a textual or tokenized description of the portion of the map in a DSL (such as RTL).

[0113] RTL can use S-expression syntax as a way to represent map information. In S-expressions, information can be grouped (such as in parentheses), where each group includes one or more items, which can be simple pieces of information such as numbers or text, or another group of S-expressions. This allows maps to be represented in a hierarchical and combinatorial manner that is relatively simple to parse. Other graphical representations can also be adopted and used at certain stages of the system to facilitate specific data operations when appropriate.

[0114] One use case for RTL is interfacing with an LLM (or other language model type), where the grammar validity of the output of the LLM can be ensured. Using a grammar to formally express data can make it easier to assert its validity compared to managing arbitrary strings or relying on a common format like JavaScript Object Notation (JSON) or another markup language (YAML), which may result in loss of semantic information. Therefore, using a formal grammar to represent the input / output of a language model can improve the robustness and reliability of the system and help ensure that the processed data has appropriate semantic meaning and is well-formed.

[0115] In at least one embodiment, one of the main entities in the RTL document is a directed graph describing a portion of a road network, such as an intersection. There are multiple possible ways to encode map coordinates, and to accommodate a small, fixed-size vocabulary, a grid-based tokenization can be implemented as a decimal representation. However, alternative methods can be used, such as using a geo-hashing representation, which provides another technique for encoding geospatial locations. However, geo-hashing typically relies on a global reference system and supports varying levels of precision, which results in the need for very large vocabularies.

[0116] Figure 5A An example graphical view 500 of an environment corresponding to a portion of a multi-lane road is shown in accordance with at least one embodiment. In order to express a map of such an environment using a language such as RTL, semantic data about the map may be used. Figure 5B An example architecture 510 for training and deployment of LLMs according to some embodiments of the present disclosure is shown. It should be understood that this arrangement and other arrangements described herein are set forth by way of example only. Other arrangements and elements (e.g., machines, interfaces, functions, commands, functional groupings, etc.) may be used in addition to or in place of those shown, and some elements may be omitted entirely. In addition, many of the elements described herein are functional entities that may be implemented as discrete or distributed components, or in combination with other components and in any suitable combination and location. The various functions described herein as being performed by an entity may be performed by hardware, firmware, and / or software. For example, the various functions may be implemented by a processor executing instructions stored in a memory. In some embodiments, the same may be used. Figures 16A-16D Example of autonomous vehicle 1600, Fig. 9 The systems, methods, and processes described herein may be performed using components, features, and / or functionality similar to the components, features, and / or functionality of the example computing device 900 and / or data center.

[0117] like Figure 5BAs shown in the example architecture 510 of , semantic information may be available (e.g., encoded in one or more maps from a database 512 or repository of truth data) and may be used to describe the map in a relevant language. In some embodiments, existing map data may be used to perform tasks such as encoding landmark features and other aspects of a map or graphic. For example, an HD map (which, in an embodiment, may be represented using an occupancy map generated from any type of sensor data (such as image data, lidar data, radar data, etc.)) may have various layers, such as a planning layer, a semantic data layer, a sensor-specific layer (e.g., a RADAR layer, a LiDAR layer, a camera layer, etc.), and / or other layer types. In order to convert the map representation to a language-based representation, such as a sequential tokenized text string, one or more of these layers may be used. An automatic conversion tool may be used to read map data in a map format and convert the data to RTL, or otherwise generate a language-based representation of the map data so that the LLM can understand and process the data. In at least one embodiment, this may include using a corpus generation component 514 that may generate a text-based representation based on an encoding or embedding from the map data, and perform training before providing the text representation to the LLM 516. Thus, the RTL may include or comprise a language that may express the topology of a road network (and / or other networks, such as networks associated with warehouses, buildings, outdoor spaces, waterways, etc.) derived from, for example, an HD map. The RTL may then be used as an interface between the LLM 416 and the HD map so that one or more LLMs may learn the underlying structure of the road network (and / or other networks).

[0118] In some embodiments, the RTL may be encoded into the HD map, for example, into one or more layers (such as a semantic layer) of the HD map. The encoded information may then be used as a training corpus to train the LLM 516. In some embodiments, features included or represented in the RTL may include, but are not limited to, traffic signs, traffic signals, poles, lane dividers, road boundaries, road markings, stop lines, waiting elements, lane elements, and / or the like. For each landmark or feature, various types of data may be represented, such as a landmark ID (e.g., a globally unique ID), a landmark number (e.g., a total number of current landmark types), spatial information (e.g., 3D latitude and longitude, size, orientation, attitude, etc.), and / or semantic information. In some embodiments, the spatial information may include 2D coordinates (which may be derived from 3D coordinates), 3D coordinates (which may be determined from 2D coordinates), 4D coordinates (which may vary over time or have a time component), boundary shape positions, and / or curve positions. Boundary shapes may be represented using float[3] for position, float[3] for size, and float[3] or [4] for orientation or attitude. Curve landmarks can be represented using a list of keypoints and their 2D or 3D locations, or using curve parameters in a parametric form. Semantic information may include, for example, but not limited to, landmark type, landmark association (e.g., associated lane ID for a traffic light, lane boundary segment location), and / or text information (e.g., text displayed on a sign).

[0119] An LLM 516 trained with an RTL (or other DSL) corpus built from a database 512 of map truth data can be queried to correct features output from a machine learning (ML) automation pipeline. The output of the LLM 516 (such as by using a writer and / or parser component or module 520) can be mapped back to the extracted features. Then, for example, a differential (e.g., diff) operation can be performed on the inferred landmarks from the automation component 518 to perform any appropriate corrections to generate a map graphic 522. Example use cases are used to infer road topology (e.g., edges) from an incomplete set of nodes (e.g., landmarks) with potential applications such as, but not limited to, tools, quality assurance (QA), and automation. In some embodiments, document embeddings can be indexed in a vector database or n-dimensional latent space (where n can represent the number of extracted features or feature types), which can then be used to cluster similar intersections, allowing unsupervised annotation and retrieval of operational design domains (ODDs) (e.g., features or landmarks).

[0120] In at least one embodiment, a structured language such as RTL may rely on the existence of a rich database of features and a graph describing the relationships of these features within a map. The methods described herein allow deterministic conversion from a set of landmark features to an RTL representation and vice versa. The graph in this example may be considered a knowledge graph, rather than general knowledge, that expresses other aspects of a road network topology or related environment. In one or more embodiments, there may be strong constraints on how coordinates are encoded so that they can be understood by a language model (LM). Location-related aspects such as coordinate, latitude, longitude, and altitude coordinate triplets may be represented using encoded values ​​or representations, such as, as a non-limiting example, a set of two characters in an alphabetical arrangement: "ab ac ad...", up to size 256.

[0121] In at least one embodiment, to tokenize or encode coordinates, for example, one, some, or all landmark features in a document may be considered, and values ​​such as their centroids may be used as the origin in an East, North, Up (ENU) coordinate system (or another coordinate system), with the altitude set to the average altitude of the coordinates in the document. For example, a radius R may be considered, such as 350 meters around the origin, divided into a grid (e.g., a 65536 x 65536 grid). Tokenization accuracy may be a function of this width, as the cells in the grid may be the smallest addressable (or indexable) cell, e.g., equal to 0.0107 m in the proposed range. Such a fixed-size grid may allow coordinates to be represented at the same scale across documents (e.g., vs. normalized relative to the bounding shape of each document).

[0122] In at least one embodiment, important entities in a language-based representation (such as an RTL document) can be implemented as a directed graph that describes a portion of an environment (such as a road network). Such a graph can be used to express connectivity between road features (topology) and can be similar to a knowledge graph. The graph nodes can then correspond to landmark features and can be typed or classified with their landmark types. The edges of the graph can correspond to relationships between road features. In some embodiments, all source nodes may include lane elements. Individual lane elements (laneEl) can be converted into small graphs, and the document can contain all nodes associated with a given laneEl. A road graph can be represented by listing all of its nodes and edges. The ordering of nodes and edges can be arbitrary; however, edges can reference nodes by their indexes in the document.

[0123] In the graph traversal representation, the edge sequence can be used to express the path on the underlying graph of the map. Such a path can include a list of laneEl nodes visited and their attributes. The attributes of the laneEl node can include its inherent attributes (such as the drivable direction of laneEl) and the attributes of the nodes it can reach (for example, the signals that this laneEl can see, and its adjacent laneEl). In this way, the RTL document may not capture the entire graph of the map, but rather the possible paths on the map. Sentences in natural language can also be considered as paths on the underlying graph of natural language. At each word, there may be many different possibilities for what the next word will be, and these possibilities can form a graph in which a particular sentence consists of a series of choices of different edges at the nodes of the graph. Figure 5C An example diagram 530 is shown that is similar to a sentence diagram, which may break down objects or tokens, and may associate additional information with the appropriate tokens or objects.

[0124] In at least one embodiment, a language model can analyze multiple sentences and determine the next word so that it is consistent with the previously arrived word. Essentially, the LLM has learned the underlying graph structure so that it can walk on the graph to generate reasonable sentences. By providing a sufficient number of potential paths, the map LLM can learn the graph of the map and can generate reasonable paths on the map. For example, when seeing a turn signal in an input sequence, the LLM can predict the turn lane for the next sign in the sequence.

[0125] Various methods can be used to encode map coordinates. For example, in various embodiments, decimal representation, grid-based tokenization, geohashing, and / or other methods can be used. To ensure a small fixed-size vocabulary, some embodiments use a grid-based tokenization method, as decimal representation and geohashing may require a larger vocabulary. Since LLM can handle sequences well, a delta-based encoding can be used to express coordinates. As an alternative, a global coordinate system can be implemented; however, while easier to parse and encode, a global coordinate system may result in sparse labeling and inefficient topology learning. It is worth noting that the cumulative absolute error is not considered problematic at this point because the length of the traversal is short and approximate relative positions may be appropriate.

[0126] In the case where each node has at least four keypoints, the first keypoint can be used as an anchor point for subsequent points in that node. As a non-limiting example, 26 can then be laid out 3 x 26 3 grid, and centered around the anchor point, which allows for an indexable area of ​​878m x 878m with an accuracy of 5cm. Figure 5E An example architecture 560 is shown that can be used to determine an output state 570. To tokenize coordinates (such as coordinates received as input in a matrix 564), all landmark features in the document can be considered, such as those that can be received as a set of tokens or other structural inputs 562. For example, a value such as a centroid can be used as the origin in the ENU coordinate system, where the altitude is set to the average altitude of the coordinates in the document. These tokens can be processed using an embedding module 566 to determine appropriate embeddings, and the coordinate inputs can be processed using an MLP 568, for example, to generate an appropriate output state 570. In at least one embodiment, a graph traversal can be generated using a random-walk on a laneEl graph. The connections of laneEl (e.g., from_laneEl and to_laneEl fields) can define all possible ways to navigate on the map. To generate a traversal, one method is to start from laneEl and recursively follow the successor laneEl to generate a path. When a branch point is encountered where there are multiple successor points, for example, the method can be to randomly select one of the successor nodes and walk along the path until the maximum token limit is reached.

[0127] Fig. 5F An example image 575 of an intersection in an example map is shown. As shown, the white arrows indicate possible traffic directions, and the two highlighted lanes are the two successor lanes of the laneEl above it. When generating a traversal, one method is to start from the top laneEl, select one of the successor laneEl to add to the path, and follow one of its successors, and so on. In order to generate a graph traversal, a random walk can be performed on the laneEl graph. In practice, a graph traversal can be generated with tasks such as:

[0128] -Task 1: Extract all laneEls in the map and put their ids into a vector.

[0129] - Task 2: Randomly select a starting laneEl from the vector, where the to_laneEl field is the possible successor laneEl that this laneEl can go to. Randomly select a successor laneEl and follow its successor laneEl, and so on.

[0130] -Task 3: When at laneEl in the path, extract the landmark features reachable from laneEl and add them to the traversal results. Once the traversal reaches the maximum landmark limit, stop the path.

[0131] In this example, nodes correspond to landmarks. A common node type in this example is "lane_el", but also includes "road_boundary", "lane_divider", "signal", "sign", "stop_line", etc. Edges can represent relationships between landmarks, and each edge has a type, such as "from_lane", "to_lane", "visible_sign", "sign_edge", etc.

[0132] According to one or more embodiments, the grammar may include a simple directed graph data structure with an arbitrary number of edges. Node attributes may be specified in the node entity and depend on the node type, which is the same as the landmark feature type. RTL documents may be generated by using logic such as the following: features are within a predefined area (e.g., 700mx 700m) defined relative to the coordinate encoding of this document; documents are written for each laneEl, the documents include relevant features as nodes and their relationships encoded as edges; and the edges of the nodes are also included. The language can be compiled using command line tools. Such tools can verify (lint) for example, check whether all nodes referenced exist in the document, and compile them to other targets, such as keyhole markup language (KML) or images, which is useful for debugging. The RTL API may provide a public interface that allows developers to generate structured language documents and also allows users to query trained models. As described herein, a schema may be provided to tokenize edge sequence paths, where the attributes of a node are represented by a single tag for illustration purposes. Although in reality, all node attributes may not fit into a single tag, there is still room to compress them into as few tags as possible.

[0133] In some embodiments, high-level contextual information can be added as a prefix to the RTL document. The generative part of the language model can learn to make RTL diagrams from the contextual information. Using this method can allow conversion from natural language descriptions, other simpler map representations, or images (camera or BEV) to languages ​​such as RTL. For example, the LLM can be described in natural language as prefixes such as "2 lane road", "4-way intersection with 2 lanes in each direction", or "large intersection with traffic lights". This high-level context description can be derived from HD and / or SD map information. In some embodiments, a model (e.g., DNN) can be used to process the image, and a description of the image or the road topology or contextual information represented in the image can be output by the model. This contextual information or semantic information can also be used to generate RTL, which can allow (using overlay capabilities) to fit the RTL to the image. Such a process can be performed using a top view or a bird's eye view (BEV) image. Such a process can also provide the ability to process an image (in a camera or BEV) into RTL (or another DSL).

[0134] Therefore, context information can be provided or preset to the query or prompt of LLM in a format different from DSL or RTL to indicate the output of generating LLM. For example, in order to understand both natural language and DSL, LLM configured for map and LLM for natural language, or LLM trained to process these two types of languages ​​can be used. As described herein, context can include a higher level description of the scene "two-lane road", "4-way intersection with 2 lanes in each direction" or "large intersection with traffic lights". In some embodiments, a coded representation of a standard or lower definition map (relative to a high-definition map) such as a navigation map can be generated and used as a prefix or pre-part of the prompt to LLM. For example, the map information can be encoded in natural language, or it can be encoded in another format that LLM can understand. For example, in an embodiment, this type of information with more or less details can be processed by one or more LLMs to generate a representation of a scene or scene described in RTL. In some embodiments, observed geometric shapes or image features can also be included in the prompt or preset to the prompt. Some of this information can be retrieved or generated using a map. For example, a map may include information encoded therein, such as "Type: intersection is asymmetric {incomplete} (NE=>SW) at t-junction," and this information may be used as part of the prompt. In some embodiments, information or descriptions, such as (number of lanes, number of lanes entering and exiting the intersection, number of turning lanes, etc.) may be retrieved or obtained by reading the map. To capture this information, in an embodiment, the RTL vocabulary may be extended to include these types of descriptions as prompt tokens. For example, an LLM may be queried with or without RTL graph information. This information may also be generated from an Open Street Map (OSM) or another map source or database, and / or from processing images or other sensor data types.

[0135] As previously mentioned, such representations of the environment can be used to perform specific tasks such as those related to simulation or testing. In at least one embodiment, the representation generated using the method proposed herein can be realistic enough to allow the planning control system of an autonomous vehicle to operate in the environment. Maintaining semantic information and building according to the understood relationships can provide a more accurate and comprehensive representation of the environment, otherwise the representation of the environment can be constructed using low-level primitive representations (such as points, lines, line segments, curves or polygons representing the shape and position of objects in the environment), and in the early stages of the reconstruction process, additional information (such as semantic data) will be discarded as in previous methods. As previously mentioned, the topological information obtained from such low-level primitives can also be fragmented due to occlusion and other such factors. The example representation generated according to at least one embodiment can retain such additional information and use one or more language models to complete the segmented representation. In addition, the representation can be further improved by using multiple perception and / or positioning modules, which can analyze different types of inputs and fuse these inputs to generate more accurate features and relationships. Where available, previous maps or representation information can also be used. The language model can be further used to fuse these types of information to generate a single aligned representation.

[0136] Figure 5GAn example process 580 of generating a textual representation of an environment that can be performed according to at least one embodiment is shown. It should be understood that for this process and other processes presented herein, unless otherwise specifically stated, additional, fewer or alternative operations or tasks may be performed or performed in a similar or alternative order, or at least partially performed in parallel within the scope of various embodiments. In addition, although this and other examples will be discussed herein with respect to maps and navigation environments, other types of environments and representations may also be used and / or generated within the scope of various embodiments. In this example process 580, input data corresponding to a set of observations of the environment may be provided 582 as input to a language model (such as a trained LLM). The set of observations may correspond to raw sensor data, features extracted from raw sensor data, feature vectors or embedding sets, or object graphs, among other such options. The language model may be a generative model trained using spatial and semantic information of various environments, and is capable of reasoning about relationships between various objects or features in the environment. A text string may be generated and received 584 as output from the language model. The text string in this example can be a single tokenized text string that includes a tokenized description of the environment, which depends in part on aspects and relationships determined or inferred for a set of observations, including semantic, geometric, and topological aspects of these observations. The text string can include information that can provide additional information about the tags in the string, which in turn can correspond to objects in the environment. In some embodiments, multiple text strings or text objects can be provided for representing the environment. The generated text string can be used to generate multiple different representations of the environment in different forms or for different purposes, and can be used to locate similar environments based at least in part on the tags included in the tokenized string.

[0137] You can use such as Figure 5G The process described above can be used to generate reconstructions of physical environments, or to generate synthesis of new environments that conform to the rules of the real world. Figure 5H, to generate a reconstruction of the physical environment based at least in part on sensor data (or other such data) captured or obtained for the physical environment. In this example process 586, sensor data of objects in the environment are captured 588. The sensor data can be captured using any appropriate sensor or device (such as those discussed herein, such as LIDAR, RADAR, camera or acoustic sensor, and other such options). Additional data of the environment can also be obtained because it can include previous map data, feature vectors, or representative text strings. The sensor data (and any additional data) can be analyzed 590 to generate a set of feature vectors corresponding to the object, such as it can correspond to an embedding in a latent space. In this example process 586, the feature vector is provided 592 as input to a trained language model, but in other embodiments, the language model can take sensor data as input and generate feature vectors or embeddings and other such options internally. (In addition to natural language information) The language model can also be trained using semantic, geometric and / or topological data of various environments and can understand the relationship between objects in the environment and the real-world rules applied to these objects. A tokenized text string (or other text representation) may be received 594 as an output of a trained language model. The tokenized text string may include a sequence of tags corresponding to objects in the environment, thereby providing semantic (and other) information about these objects. In at least some embodiments, the tag sequence may be used as a flattened map graphic and may be used to generate a realistic environment. In this example, the tokenized text string representing the spatial and semantic information of the environment may be provided 596 to a reconstruction generator (which may also include a generative machine learning model) to generate a digital reconstruction of the environment, such as may correspond to a bird's eye view map, an HD map, or a 3D virtual environment, among other such options.

[0138] Aspects of the various methods presented herein may be sufficiently lightweight to be performed in real time in a variety of locations (such as on a device), such as a client device including a personal computer or a game console. Such processing may be performed on or for content generated on or received by a client device or received from an external source, such as streaming data or other content received from a cloud server 620 or a third-party service 660 over at least one network, among other such options. In some cases, at least a portion of the processing, generation, synthesis, and / or determination of the content may be performed by one of these other devices, systems, or entities, which is then provided to the client device (or another such recipient) for presentation or another such use.

[0139] As an example, Figure 6An example network configuration 600 is shown that can be used to provide, generate, modify, encode, process, fuse and / or transmit generated data or other such content. In at least one embodiment, a client device 602 can generate or receive data for a session using components of a content application 604 on the client device 602 and data stored locally on the client device. In at least one embodiment, a content application 624 executed on a computer or processor 620 (e.g., a cloud server or control system) can initiate a session associated with at least one client device 602 (e.g., a vehicle or robot), such as using a session manager and user data stored in a user database 636, and can cause content such as one or more digital assets (e.g., implicit and / or explicit object representations or maps) from an asset library 634 to be determined by a content manager 626. The content manager 626 can work with a trained language module 628 to generate a text-based representation of an environment based on several types of input data. This can include maps and / or sensor data, where the map data can be generated and / or recognized by a map module 630, and the map data can be associated with perception data generated by a perception module 632 based on captured sensor data or other such information. At least a portion of the generated text-based representation may be transmitted to the client device 602 using an appropriate transmission manager 622 for transmission via download, streaming, or other such transmission channels. An encoder may be used to encode and / or compress at least some of such data before transmission to the client device 602. In at least one embodiment, the client device 602 receiving such content may provide the content to a corresponding content application 604, which may also or alternatively include a graphical user interface 610 and a content manager 612 for providing, synthesizing, rendering, composing, modifying, or using content for presentation (or other purposes) on or by the client device 602. The content application 604 may also include a language module 614 that may perform various generation tasks (such as updating or enhancing a text-based representation, or using a map and / or sensory data on the client device itself to generate a text description). In some embodiments, the computer 620 and the client device 602 are able to communicate directly without transmitting data over a network 640 to avoid issues such as latency and availability. The decoder can also be used to decode data received over the network 640 for presentation via the client device 602, such as images or video content received via a display device 606, and audio (such as sounds and music) received via at least one audio playback device 608 (e.g., a speaker or headphones).In at least one embodiment, at least some of the content may already be stored on, rendered on, or accessible by the client device 602, such that at least such portions of the content do not need to be transmitted over the network 640, such as where such content may have previously been downloaded or stored locally on a hard disk or optical disk. In at least one embodiment, a transmission mechanism such as a data stream may be used to transmit the content from the cloud server 620 or user database 636 to the client device 602. In at least one embodiment, at least a portion of the content may be obtained, enhanced, and / or streamed from another source such as a third-party service 660 or other client device 650, which may also include a content application 662 for generating, enhancing, or providing content. In at least one embodiment, portions of the functionality may be performed using multiple computing devices or multiple processors within one or more computing devices, such as, for example, the multiple processors may include a combination of a CPU and a GPU (graphics processing unit).

[0140] In this example, these client devices may include any appropriate computing device, such as a desktop computer, a notebook computer, a set-top box, a streaming device, a game console, a smart phone, a tablet computer, a VR helmet, an AR goggles, a wearable computer or a smart TV. Each client device can submit a request across at least one wired or wireless network, which may include the Internet, an Ethernet, a local area network (LAN), or a cellular network, and other such options. In such an example, these requests may be submitted to an address associated with a cloud provider, and the cloud provider may operate or control one or more electronic resources in a cloud provider environment, such as may include a data center or a server farm. In at least one embodiment, such a request may be received or processed by at least one edge server located on the edge of the network and outside of at least one security layer associated with the cloud provider environment. In this way, latency can be reduced by allowing client devices to interact with closer servers while also improving the security of resources in the cloud provider environment.

[0141] In at least one embodiment, such a system can be used to perform graphics rendering operations. In other embodiments, such a system can be used for other purposes, such as for providing image or video content to test or verify autonomous machine applications, or for performing deep learning operations. In at least one embodiment, such a system can be implemented using an edge device, or can include one or more virtual machines (VMs). In at least one embodiment, such a system can be implemented at least in part in a data center or at least in part using cloud computing resources.

[0142] Reasoning and training logic

[0143] Fig. 7A Inference and / or training logic 715 is shown for performing reasoning and / or training operations associated with one or more embodiments. Fig. 7A and / or Figure 7B Details regarding the inference and / or training logic 715 are provided.

[0144] In at least one embodiment, the inference and / or training logic 715 may include, but is not limited to, code and / or data storage 701 for storing forward and / or output weights and / or input / output data, and / or other parameters for configuring neurons or layers of a neural network trained and / or used for inference in aspects of one or more embodiments. In at least one embodiment, the training logic 715 may include or be coupled to code and / or data storage 701 for storing graph code or other software to control timing and / or sequence, where weights and / or other parameter information is loaded to configure logic, including integer and / or floating point units (collectively referred to as arithmetic logic units (ALUs)). In at least one embodiment, the code (such as graph code) loads weights or other parameter information into a processor ALU based on the architecture of the neural network to which the code corresponds. In at least one embodiment, the code and / or data storage 701 stores input / output data during training and / or inference using aspects of one or more embodiments and / or weight parameters during forward propagation of each layer of a neural network trained or used in conjunction with one or more embodiments. In at least one embodiment, any portion of code and / or data storage 701 may be included within other on-chip or off-chip data storage, including a processor's L1, L2, or L3 cache or system memory.

[0145] In at least one embodiment, any portion of code and / or data storage 701 may be internal or external to one or more processors or other hardware logic devices or circuits. In at least one embodiment, code and / or data storage 701 may be cache memory, dynamic random addressable memory ("DRAM"), static random addressable memory ("SRAM"), non-volatile memory (e.g., flash memory), or other storage. In at least one embodiment, the choice of whether code and / or data storage 701 is internal or external to a processor, for example, or consists of DRAM, SRAM, flash memory, or some other storage type, may depend on the available storage space on-chip or off-chip for storage, the latency requirements of the training and / or inference functions being performed, the batch size of the data used in inference and / or training of the neural network, or some combination of these factors.

[0146] In at least one embodiment, the inference and / or training logic 715 may include, but is not limited to, code and / or data storage 705 to store the backpropagation and / or output weights and / or input / output data corresponding to the neurons or layers of the neural network trained and / or used for inference in aspects of one or more embodiments. In at least one embodiment, during training and / or inference using aspects of one or more embodiments, the code and / or data storage 705 stores the weight parameters and / or input / output data for each layer of the neural network trained or used in conjunction with one or more embodiments during backpropagation of the input / output data and / or weight parameters. In at least one embodiment, the training logic 715 may include or be coupled to code and / or data storage 705 for storing graph code or other software to control timing and / or sequence, wherein the weights and / or other parameter information is loaded to configure logic, which includes integer and / or floating point units (collectively referred to as arithmetic logic units (ALUs)). In at least one embodiment, the code (such as graph code) loads the weights or other parameter information into the processor ALU based on the architecture of the neural network to which the code corresponds. In at least one embodiment, any portion of code and / or data storage 705 may be included with other on-chip or off-chip data storage, including the processor's L1, L2, or L3 cache or system memory. In at least one embodiment, any portion of code and / or data storage 705 may be internal or external to one or more processors or other hardware logic devices or circuits. In at least one embodiment, code and / or data storage 705 may be cache memory, DRAM, SRAM, non-volatile memory (e.g., flash memory), or other storage. In at least one embodiment, the choice of whether code and / or data storage 705 is internal or external to the processor, for example, whether it is composed of DRAM, SRAM, flash memory, or some other storage type, depends on whether the available storage is on-chip or off-chip, the latency requirements of the training and / or inference functions being performed, the data batch size used in the inference and / or training of the neural network, or some combination of these factors.

[0147] In at least one embodiment, code and / or data storage 701 and code and / or data storage 705 may be separate storage structures. In at least one embodiment, code and / or data storage 701 and code and / or data storage 705 may be the same storage structure. In at least one embodiment, code and / or data storage 701 and code and / or data storage 705 may be partially the same storage structure and partially separate storage structures. In at least one embodiment, any portion of code and / or data storage 701 and code and / or data storage 705 may be included with other on-chip or off-chip data storage, including the processor's L1, L2, or L3 cache or system memory.

[0148] In at least one embodiment, the inference and / or training logic 715 may include, but is not limited to, one or more arithmetic logic units (“ALUs”) 710 (including integer and / or floating point units) for performing logical and / or mathematical operations based at least in part on or as directed by training and / or inference code (e.g., graph code), the results of which may produce activations (e.g., output values ​​from a layer or neuron within a neural network) stored in activation storage 720, which is a function of input / output and / or weight parameter data stored in code and / or data storage 701 and / or code and / or data storage 705. In at least one embodiment, activations are performed in response to executing instructions or other code, linear algebra and / or matrix-based mathematics performed by ALU 710 to generate activations stored in activation storage 720, where weight values ​​stored in code and / or data storage 701 and / or code and / or data storage 705 are used as operands with other values, such as bias values, gradient information, momentum values, or other parameters or hyperparameters, any or all of which may be stored in code and / or data storage 701 or code and / or data storage 705 or other on-chip or off-chip storage.

[0149] In at least one embodiment, one or more ALUs 710 are included in one or more processors or other hardware logic devices or circuits, while in another embodiment, one or more ALUs 710 may be outside a processor or other hardware logic device or circuit that uses them (e.g., a coprocessor). In at least one embodiment, one or more ALUs 710 may be included within an execution unit of a processor, or otherwise included in a group of ALUs accessible by an execution unit of a processor, which may be within the same processor or distributed between different processors of different types (e.g., a central processing unit, a graphics processing unit, a fixed function unit, etc.). In at least one embodiment, code and / or data storage 701, code and / or data storage 705, and activation storage 720 may be the same processor or other hardware logic device or circuit, while in another embodiment, they may be in different processors or other hardware logic devices or circuits or some combination of the same and different processors or other hardware logic devices or circuits. In at least one embodiment, any portion of activation storage 720 may be included with other on-chip or off-chip data storage, including the processor's L1, L2, or L3 cache or system memory. Additionally, inference and / or training code may be stored with other code accessible to a processor or other hardware logic or circuitry and may be retrieved and / or processed using the processor's fetch, decode, schedule, execute, exit, and / or other logic circuitry.

[0150] In at least one embodiment, activation storage 720 may be cache memory, DRAM, SRAM, non-volatile memory (e.g., flash memory), or other storage. In at least one embodiment, activation storage 720 may be completely or partially internal or external to one or more processors or other logic circuits. In at least one embodiment, activation storage 720 may be selected to be internal or external to a processor, for example, or include DRAM, SRAM, flash memory, or other storage types, depending on the storage available on-chip or off-chip, the latency requirements for performing training and / or inference functions, the batch size of data used in inferencing and / or training neural networks, or some combination of these factors. In at least one embodiment, Fig. 7A The inference and / or training logic 715 shown in FIG. 7 may be used in conjunction with an application specific integrated circuit (“ASIC”), such as the ASIC from Google. Processing unit from Graphcore TM Inference Processing Unit (IPU) from Intel Corp (e.g., "Lake Crest") processor. In at least one embodiment, Fig. 7A The illustrated inference and / or training logic 715 may be used in conjunction with central processing unit (“CPU”) hardware, graphics processing unit (“GPU”) hardware, or other hardware such as a field programmable gate array (“FPGA”).

[0151] Figure 7B Inference and / or training logic 715 is shown in accordance with at least one or more embodiments. In at least one embodiment, the reasoning and / or training logic 715 may include, but is not limited to, hardware logic in which computing resources are dedicated or otherwise uniquely used in conjunction with weight values ​​or other information corresponding to one or more layers of neurons within a neural network. In at least one embodiment, Figure 7B The inference and / or training logic 715 shown in FIG. 7 may be used in conjunction with an application specific integrated circuit (ASIC), such as the ASIC from Google. Processing unit from Graphcore TM Inference Processing Unit (IPU) from Intel Corp (e.g., "Lake Crest") processor. In at least one embodiment, Figure 7BThe inference and / or training logic 715 shown in can be used in conjunction with central processing unit (CPU) hardware, graphics processing unit (GPU) hardware, or other hardware (e.g., field programmable gate array (FPGA)). In at least one embodiment, the inference and / or training logic 715 includes, but is not limited to, code and / or data storage 701 and code and / or data storage 705, which can be used to store code (e.g., graph code), weight values, and / or other information, including bias values, gradient information, momentum values, and / or other parameter or hyperparameter information. Figure 7B In at least one embodiment shown in , each of code and / or data storage 701 and code and / or data storage 705 is associated with a dedicated computing resource (e.g., computing hardware 702 and computing hardware 706), respectively. In at least one embodiment, each of computing hardware 702 and computing hardware 706 includes one or more ALUs that perform mathematical functions (e.g., linear algebraic functions) solely on the information stored in code and / or data storage 701 and code and / or data storage 705, respectively, and the results of performing the functions are stored in activation storage 720.

[0152] In at least one embodiment, each of the code and / or data stores 701 and 705 and the corresponding computing hardware 702 and 706 corresponds to a different layer of the neural network, such that activations from one "storage / compute pair 701 / 702" of the code and / or data store 701 and computing hardware 702 are provided as inputs to the next "storage / compute pair 705 / 706" of the code and / or data store 705 and computing hardware 706, so as to reflect the conceptual organization of the neural network. In at least one embodiment, each storage / compute pair 701 / 702 and 705 / 706 can correspond to more than one neural network layer. In at least one embodiment, additional storage / compute pairs (not shown) can be included in the inference and / or training logic 715 after or in parallel with the storage / compute pairs 701 / 702 and 705 / 706.

[0153] Data Center

[0154] Figure 8 An example data center 800 is shown in which at least one embodiment may be used. In at least one embodiment, the data center 800 includes a data center infrastructure layer 810, a framework layer 820, a software layer 830, and an application layer 840.

[0155] In at least one embodiment, Figure 8As shown, the data center infrastructure layer 810 may include a resource coordinator 812, group computing resources 814, and node computing resources ("node CRs") 816 (1)-816 (N), where "N" represents any positive integer. In at least one embodiment, the node CRs 816 (1)-816 (N) may include, but are not limited to, any number of central processing units ("CPUs") or other processors (including accelerators, field programmable gate arrays (FPGAs), graphics processors, etc.), memory devices (e.g., dynamic read-only memories), storage devices (e.g., solid-state drives or disk drives), network input / output ("NW I / O") devices, network switches, virtual machines ("VMs"), power modules and cooling modules, etc. In at least one embodiment, one or more of the node CRs 816 (1)-816 (N) may be a server having one or more of the above computing resources.

[0156] In at least one embodiment, the grouped computing resources 814 may include a separate grouping (not shown) of node CRs housed in one or more racks, or many racks (also not shown) housed in data centers at various geographic locations. The separate grouping of node CRs within the grouped computing resources 814 may include computing, networks, memory, or storage resources that can be configured or allocated to support groupings of one or more workloads. In at least one embodiment, several node CRs including a CPU or processor may be grouped in one or more racks to provide computing resources to support one or more workloads. In at least one embodiment, one or more racks may also include any number of power modules, cooling modules, and network switches in any combination.

[0157] In at least one embodiment, resource coordinator 812 may configure or otherwise control one or more node CRs 816(1)-816(N) and / or grouped computing resources 814. In at least one embodiment, resource coordinator 812 may include a software design infrastructure ("SDI") management entity for data center 800. In at least one embodiment, resource coordinator 812 may include hardware, software, or some combination thereof.

[0158] In at least one embodiment, Figure 8As shown, the framework layer 820 includes a job scheduler 822, a configuration manager 824, a resource manager 826, and a distributed file system 828. In at least one embodiment, the framework layer 820 may include a framework that supports software 832 of the software layer 830 and / or one or more applications 842 of the application layer 840. In at least one embodiment, the software 832 or the application 842 may include web-based service software or applications, such as services or applications provided by Amazon WebServices, Google Cloud, and Microsoft Azure. In at least one embodiment, the framework layer 820 may be, but is not limited to, a free and open source software network application framework, such as Apache SparkTM (hereinafter referred to as "Spark") that can use the distributed file system 828 for large-scale data processing (e.g., "big data"). In at least one embodiment, the job scheduler 822 may include a Spark driver to facilitate scheduling of workloads supported by various layers of the data center 800. In at least one embodiment, the configuration manager 824 may be able to configure different layers, such as the software layer 830 and the framework layer 820 including Spark and a distributed file system 828 for supporting large-scale data processing. In at least one embodiment, the resource manager 826 can manage cluster or group computing resources mapped to or allocated to support the distributed file system 828 and the job scheduler 822. In at least one embodiment, the cluster or group computing resources can include group computing resources 814 on the data center infrastructure layer 810. In at least one embodiment, the resource manager 826 can coordinate with the resource coordinator 812 to manage these mapped or allocated computing resources.

[0159] In at least one embodiment, the software 832 included in the software layer 830 may include software used by at least a portion of the node CRs 816(1)-816(N), the grouped computing resources 814, and / or the distributed file system 828 of the framework layer 820. The one or more types of software may include, but are not limited to, Internet web page search software, email virus scanning software, database software, and streaming video content software.

[0160] In at least one embodiment, one or more applications 842 included in the application layer 840 may include one or more types of applications used by at least a portion of the node CRs 816(1)-816(N), the grouped computing resources 814, and / or the distributed file system 828 of the framework layer 820. The one or more types of applications may include, but are not limited to, any number of genomics applications, cognitive computing, and machine learning applications, including training or inference software, machine learning framework software (e.g., PyTorch, TensorFlow, Caffe, etc.), or other machine learning applications used in conjunction with one or more embodiments.

[0161] In at least one embodiment, any of the configuration manager 824, resource manager 826, and resource coordinator 812 can implement any number and type of self-modification actions based on any number and type of data acquired in any technically feasible manner. In at least one embodiment, the self-modification actions can relieve a data center operator of the data center 800 from making potentially bad configuration decisions and can avoid underutilized and / or poorly performing portions of the data center.

[0162] In at least one embodiment, the data center 800 may include tools, services, software, or other resources to train one or more machine learning models or use one or more machine learning models to predict or infer information according to one or more embodiments described herein. For example, in at least one embodiment, a machine learning model may be trained by calculating weight parameters according to a neural network architecture using the software and computing resources described above with respect to the data center 800. In at least one embodiment, by using weight parameters calculated by one or more training techniques described herein, information may be inferred or predicted using trained machine learning models corresponding to one or more neural networks using the resources described above with respect to the data center 800.

[0163] In at least one embodiment, the data center can use a CPU, an application-specific integrated circuit (ASIC), a GPU, an FPGA, or other hardware to use the above resources to perform training and / or reasoning. In addition, one or more of the above software and / or hardware resources can be configured as a service to allow users to train or perform information reasoning, such as image recognition, speech recognition, or other artificial intelligence services.

[0164] The reasoning and / or training logic 715 is used to perform reasoning and / or training operations associated with one or more embodiments. Fig. 7A and / or Figure 7B Provides details about the reasoning and / or training logic 715. In at least one embodiment, the reasoning and / or training logic 715 may be implemented in the system Figure 8 for use in inference or prediction operations based at least in part on weight parameter inferences or predictions computed using neural network training operations, neural network functions and / or architectures, or neural network use cases described herein.

[0165] These components may be used to generate a single consistent tokenized description of at least a portion of a physical environment based in part on the set of observations and the aligned map data.

[0166] Computer Systems

[0167] Fig. 9 900, which may be a system of interconnected devices and components, a system on a chip (SOC), or some combination thereof formed with a processor, which may include an execution unit to execute instructions. In at least one embodiment, in accordance with the present disclosure, such as the embodiments described herein, the computer system 900 may include, but is not limited to, components, such as a processor 902, whose execution unit includes logic to execute algorithms for process data. In at least one embodiment, the computer system 900 may include a processor, such as a processor available from Intel Corporation of Santa Clara, California. Processor family, XeonTM, XScaleTM and / or StrongARMTM, Core TM or Nervana TM microprocessor, although other systems (including PCs with other microprocessors, engineering workstations, set-top boxes, etc.) may also be used. In at least one embodiment, computer system 900 may execute a version of the WINDOWS operating system available from Microsoft Corporation of Redmond, Wash., although other operating systems (e.g., UNIX and Linux), embedded software, and / or graphical user interfaces may also be used.

[0168] Embodiments may be used in other devices, such as handheld devices and embedded applications. Some examples of handheld devices include cellular phones, Internet Protocol (Internet Protocol) devices, digital cameras, personal digital assistants ("PDAs"), and handheld PCs. In at least one embodiment, the embedded application may include a microcontroller, a digital signal processor ("DSP"), a system on a chip, a network computer ("NetPC"), a set-top box, a network hub, a wide area network ("WAN") switch, or any other system that can execute one or more instructions according to at least one embodiment.

[0169] In at least one embodiment, the computer system 900 may include, but is not limited to, a processor 902, which may include, but is not limited to, one or more execution units 908 to perform machine learning model training and / or reasoning according to the techniques described herein. In at least one embodiment, the computer system 900 is a single-processor desktop or server system, but in another embodiment, the computer system 900 may be a multi-processor system. In at least one embodiment, the processor 902 may include, but is not limited to, a complex instruction set computing ("CISC") microprocessor, a reduced instruction set computing ("RISC") microprocessor, a very long instruction word computing ("VLIW") microprocessor, a processor that implements an instruction set combination, or any other processor device, such as a digital signal processor. In at least one embodiment, the processor 902 may be coupled to a processor bus 910, which may transmit data signals between the processor 902 and other components in the computer system 900.

[0170] In at least one embodiment, processor 902 may include, but is not limited to, a level 1 ("L1") internal cache memory ("cache") 904. In at least one embodiment, processor 902 may have a single internal cache or multiple levels of internal cache. In at least one embodiment, cache 904 may reside external to processor 902. Other embodiments may also include a combination of internal and external caches, depending on the particular implementation and needs. In at least one embodiment, register file 906 may store different types of data in various registers, including, but not limited to, integer registers, floating point registers, status registers, and instruction pointer registers.

[0171] In at least one embodiment, an execution unit 908, including but not limited to logic to perform integer and floating point operations, is also located in the processor 902. In at least one embodiment, the processor 902 may also include a microcode ("ucode") read-only memory ("ROM") for storing microcode for certain macro instructions. In at least one embodiment, the execution unit 908 may include logic for processing a packed instruction set 909. In at least one embodiment, by including the packed instruction set 909 in the instruction set of a general-purpose processor, and the associated circuitry to execute the instructions, operations used by many multimedia applications may be performed using packed data in the general-purpose processor 902. In one or more embodiments, many multimedia applications may be accelerated and executed more efficiently by using the full width of the processor's data bus 910 to perform operations on packed data, which may not require the transfer of smaller units of data on the processor's data bus 910 to perform one or more operations one data element at a time.

[0172] In at least one embodiment, execution unit 908 may also be used in microcontrollers, embedded processors, graphics devices, DSPs, and other types of logic circuits. In at least one embodiment, computer system 900 may include, but is not limited to, memory 920. In at least one embodiment, memory 920 may be implemented as a dynamic random access memory ("DRAM") device, a static random access memory ("SRAM") device, a flash memory device, or other storage device. In at least one embodiment, memory 920 may store instructions 919 and / or data 921 represented by data signals that may be executed by processor 902.

[0173] In at least one embodiment, the system logic chip can be coupled to the processor bus 910 and the memory 920. In at least one embodiment, the system logic chip can include, but is not limited to, a memory controller hub ("MCH") 916, and the processor 902 can communicate with the MCH 916 via the processor bus 910. In at least one embodiment, the MCH 916 can provide a high-bandwidth memory path 918 to the memory 920 for instruction and data storage and for storage of graphics commands, data, and textures. In at least one embodiment, the MCH 916 can initiate data signals between the processor 902, the memory 920, and other components in the computer system 900, and bridge data signals between the processor bus 910, the memory 920, and the system I / O 922. In at least one embodiment, the system logic chip can provide a graphics port for coupling to a graphics controller. In at least one embodiment, the MCH 916 can be coupled to the memory 920 via a high-bandwidth memory path 918, and the graphics / video card 912 can be coupled to the MCH 916 via an Accelerated Graphics Port ("AGP") interconnect 914.

[0174] In at least one embodiment, the computer system 900 may use a system I / O 922, which is a proprietary hub interface bus, for coupling the MCH 916 to an I / O controller hub ("ICH") 930. In at least one embodiment, the ICH 930 may provide direct connections to certain I / O devices through a local I / O bus. In at least one embodiment, the local I / O bus may include, but is not limited to, a high-speed I / O bus for connecting peripheral devices to the memory 920, the chipset, and the processor 902. Examples may include, but are not limited to, an audio controller 929, a firmware hub ("Flash BIOS") 928, a wireless transceiver 926, a data store 924, a traditional I / O controller 923 including a user input and keyboard interface 925, a serial expansion port 927 (e.g., a universal serial bus (USB) port), and a network controller 934. The data store 924 may include a hard drive, a floppy drive, a CD-ROM device, a flash memory device, or other mass storage device.

[0175] In at least one embodiment, Fig. 9 The system is shown as comprising interconnected hardware devices or "chips", while in other embodiments, Fig. 9 An exemplary system on a chip ("SoC") may be shown. In at least one embodiment, the devices may be interconnected with a proprietary interconnect, a standardized interconnect (e.g., PCIe), or some combination thereof. In at least one embodiment, one or more components of computer system 900 are interconnected using a Compute Express Link (CXL) interconnect.

[0176] Reasoning and / or training logic 715 is used to perform reasoning and / or training operations related to one or more embodiments. Fig. 7A and / or Figure 7B Provide details about the reasoning and / or training logic 1015. In at least one embodiment, the reasoning and / or training logic 715 may be Fig. 9 for use in a system for performing inference or prediction operations based at least in part on weight parameters computed using neural network training operations, neural network functions and / or architectures, or neural network use cases described herein.

[0177] Such a component may be used to generate a single consistent tokenized description of at least a portion of a physical environment based in part on a set of observations and aligned map data.

[0178] Fig.10 1 is a block diagram illustrating an electronic device 1000 for using a processor 1010 according to at least one embodiment. In at least one embodiment, the electronic device 1000 may be, for example but not limited to, a notebook computer, a tower server, a rack server, a blade server, a laptop computer, a desktop computer, a tablet computer, a mobile device, a phone, an embedded computer, or any other suitable electronic device.

[0179] In at least one embodiment, the electronic device 1000 may include, but is not limited to, a processor 1010 communicatively coupled to any suitable number or kind of components, peripherals, modules, or devices. In at least one embodiment, the processor 1010 is coupled using a bus or interface, such as an I2C bus, a system management bus ("SMBus"), a low pin count (LPC) bus, a serial peripheral interface ("SPI"), a high-definition audio ("HDA") bus, a serial advanced technology attachment ("SATA") bus, a universal serial bus ("USB") (versions 1, 2, 3), or a universal asynchronous receiver / transmitter ("UART") bus. In at least one embodiment, Fig.10 An electronic device 1000 is shown, the system includes interconnected hardware devices or "chips", and in other embodiments, Fig.10 An exemplary system on a chip ("SoC") may be shown. In at least one embodiment, Fig.10 The devices shown in can be interconnected with a proprietary interconnect, a standardized interconnect (e.g., PCIe), or some combination thereof. In at least one embodiment, Fig.10 One or more components of the system are interconnected using Compute Express Link (CXL) interconnect lines.

[0180] In at least one embodiment, Fig.10It may include a display 1024, a touch screen 1025, a touch pad 1030, a near field communication unit (“NFC”) 1045, a sensor hub 1040, a thermal sensor 1046, a fast chipset (“EC”) 1035, a trusted platform module (“TPM”) 1038, a BIOS / firmware / flash memory (“BIOS, FW Flash”) 1022, a DSP 1060, a drive 1020 (e.g., a solid state disk (“SSD”) or a hard disk drive (“HDD”)), a wireless local area network unit (“WLAN”) 1050, a Bluetooth unit 1052, a wireless wide area network unit (“WWAN”) 1056, a global positioning system (GPS) unit 1055, a camera (“USB 3.0 camera”) 1054 (e.g., a USB 3.0 camera), and / or a low power double data rate (“LPDDR”) memory unit (“LPDDR3”) 1015 implemented with, for example, the LPDDR3 standard. These components may each be implemented in any suitable manner.

[0181] In at least one embodiment, other components may be communicatively coupled to the processor 1010 through the components discussed herein. In at least one embodiment, the accelerometer 1041, ambient light sensor (“ALS”) 1042, compass 1043, and gyroscope 1044 may be communicatively coupled to the sensor hub 1040. In at least one embodiment, the thermal sensor 1039, fan 1037, keyboard 1036, and touchpad 1030 may be communicatively coupled to the EC 1035. In at least one embodiment, the speaker 1063, earphone 1064, and microphone (“mic”) 1065 may be communicatively coupled to the audio unit (“audio codec and class D amplifier”) 1062, which in turn may be communicatively coupled to the DSP 1060. In at least one embodiment, the audio unit 1062 may include, for example, but not limited to, an audio encoder / decoder (“codec”) and a class D amplifier. In at least one embodiment, the SIM card (“SIM”) 1057 may be communicatively coupled to the WWAN unit 1056. In at least one embodiment, components such as the WLAN unit 1050 and the Bluetooth unit 1052 and the WWAN unit 1056 may be implemented as a next generation form factor (NGFF).

[0182] Reasoning and / or training logic 715 is used to perform reasoning and / or training operations associated with one or more embodiments. Fig. 7A and / or Figure 7B Provides details about the reasoning and / or training logic 1015. In at least one embodiment, the reasoning and / or training logic 715 may be in the system Fig.10for use in reasoning or predicting operations based at least in part on weight parameters computed using neural network training operations, neural network functions and / or architectures, or neural network use cases described herein.

[0183] Such a component may be used to generate a single consistent tokenized description of at least a portion of a physical environment based in part on a set of observations and aligned map data.

[0184] Fig.11 1100 is a block diagram of a processing system according to at least one embodiment. In at least one embodiment, processing system 1100 includes one or more processors 1102 and one or more graphics processors 1108, and can be a single-processor desktop system, a multi-processor workstation system, or a server system with a large number of processors 1102 or processor cores 1107. In at least one embodiment, processing system 1100 is a processing platform incorporated within a system-on-chip (SoC) integrated circuit for use in a mobile, handheld, or embedded device.

[0185] In at least one embodiment, the processing system 1100 may include or be incorporated into a server-based gaming platform, including a gaming console, a mobile gaming console, a handheld gaming console, or an online gaming console for gaming and media consoles. In at least one embodiment, the processing system 1100 is a mobile phone, a smart phone, a tablet computing device, or a mobile Internet device. In at least one embodiment, the processing system 1100 may also include a wearable device coupled to or integrated in a wearable device, such as a smart watch wearable device, a smart glasses device, an augmented reality device, or a virtual reality device. In at least one embodiment, the processing system 1100 is a television or set-top box device having one or more processors 1102 and a graphical interface generated by one or more graphics processors 1108.

[0186] In at least one embodiment, one or more processors 1102 each include one or more processor cores 1107 to process instructions that, when executed, perform operations for system and user software. In at least one embodiment, each of the one or more processor cores 1107 is configured to process a specific instruction set 1109. In at least one embodiment, the instruction set 1109 can facilitate complex instruction set computing (CISC), reduced instruction set computing (RISC), or calculate by very long instruction words (VLIW). In at least one embodiment, the processor cores 1107 can each process different instruction sets 1109, and the instruction sequence can include instructions that help emulate other instruction sets. In at least one embodiment, the processor core 1107 can also include other processing devices, such as a digital signal processor (DSP).

[0187] In at least one embodiment, the processor 1102 includes a cache memory ("cache") 1104. In at least one embodiment, the processor 1102 may have a single internal cache or multiple levels of internal cache. In at least one embodiment, the cache 1104 is shared between various components of the processor 1102. In at least one embodiment, the processor 1102 also uses an external cache (e.g., a level 3 (L3) cache or a last level cache (LLC)) (not shown), which may be shared between the processor cores 1107 using known cache alignment techniques. In at least one embodiment, the processor 1102 additionally includes a register file 1106, which may include different types of registers (e.g., integer registers, floating point registers, status registers, and instruction pointer registers) for storing different types of data. In at least one embodiment, the register file 1106 may include general purpose registers or other registers.

[0188] In at least one embodiment, one or more processors 1102 are coupled to one or more interface buses 1110 to transmit communication signals, such as address, data, or control signals, between the processor 1102 and other components in the processing system 1100. In at least one embodiment, the interface bus 1110 can be a processor bus in one embodiment, such as a version of a direct media interface (DMI) bus. In at least one embodiment, the interface bus 1110 is not limited to a DMI bus, and can include one or more peripheral component interconnect buses (e.g., PCI, PCI Express), memory buses, or other types of interface buses. In at least one embodiment, the processor 1102 includes an integrated memory controller 1116 and a platform controller hub 1130. In at least one embodiment, the memory controller 1116 facilitates communication between the memory device 1120 and other components of the processing system 1100, while the platform controller hub (PCH) 1130 provides connections to input / output (I / O) devices through a local I / O bus.

[0189] In at least one embodiment, the memory device 1120 may be a dynamic random access memory (DRAM) device, a static random access memory (SRAM) device, a flash memory device, a phase change memory device, or have appropriate performance to be used as a processor memory. In at least one embodiment, the memory device 1120 may be used as a system memory of the processing system 1100 to store data 1122 and instructions 1121 for use when one or more processors 1102 execute applications or processes. In at least one embodiment, the memory controller 1116 is also coupled to an optional external graphics processor 1112, which may communicate with one or more graphics processors 1108 in the processor 1102 to perform graphics and media operations. In at least one embodiment, a display device 1111 may be connected to the processor 1102. In at least one embodiment, the display device 1111 may include one or more of the internal display devices, such as in a mobile electronic device or laptop device or an external display device connected via a display interface (e.g., DisplayPort, etc.). In at least one embodiment, the display device 1111 may include a head mounted display (HMD), such as a stereoscopic display device used in virtual reality (VR) applications or augmented reality (AR) applications.

[0190] In at least one embodiment, the platform controller hub 1130 allows peripheral devices to be connected to the storage device 1120 and the processor 1102 via a high-speed I / O bus. In at least one embodiment, the I / O peripherals include, but are not limited to, an audio controller 1146, a network controller 1134, a firmware interface 1128, a wireless transceiver 1126, a touch sensor 1125, a data storage device 1124 (e.g., a hard drive, flash memory, etc.). In at least one embodiment, the data storage device 1124 can be connected via a storage interface (e.g., SATA) or via a peripheral bus, such as a peripheral component interconnect bus (e.g., PCI, PCIe). In at least one embodiment, the touch sensor 1125 can include a touch screen sensor, a pressure sensor, or a fingerprint sensor. In at least one embodiment, the wireless transceiver 1126 can be a Wi-Fi transceiver, a Bluetooth transceiver, or a mobile network transceiver, such as a 3G, 4G, or long-term evolution (LTE) transceiver. In at least one embodiment, the firmware interface 1128 allows communication with the system firmware and can be, for example, a unified extensible firmware interface (UEFI). In at least one embodiment, the network controller 1134 can allow network connection to a wired network. In at least one embodiment, a high-performance network controller (not shown) is coupled to the interface bus 1110. In at least one embodiment, the audio controller 1146 is a multi-channel high-definition audio controller. In at least one embodiment, the processing system 1100 includes an optional legacy I / O controller 1140 for coupling legacy (e.g., Personal System 2 (PS / 2)) devices to the system 1100. In at least one embodiment, the platform controller hub 1130 can also be connected to one or more universal serial bus (USB) controllers 1142, which connect input devices such as a keyboard and mouse 1143 combination, a camera 1144, or other USB input devices.

[0191] In at least one embodiment, instances of the memory controller 1116 and the platform controller hub 1130 may be integrated into a discrete external graphics processor, such as the external graphics processor 1112. In at least one embodiment, the platform controller hub 1130 and / or the memory controller 1116 may be external to one or more processors 1102. For example, in at least one embodiment, the processing system 1100 may include the external memory controller 1116 and the platform controller hub 1130, which may be configured as a memory controller hub and a peripheral controller hub in a system chipset that communicates with the processor 1102.

[0192] Reasoning and / or training logic 715 is used to perform reasoning and / or training operations associated with one or more embodiments. Fig. 7A and / or Figure 7BDetail is provided regarding the inference and / or training logic 715. In at least one embodiment, some or all of the inference and / or training logic 715 may be incorporated into the processing system 1100. For example, in at least one embodiment, the training and / or inference techniques described herein may use one or more ALUs embodied in a graphics processor. Additionally, in at least one embodiment, the inference and / or training operations described herein may use a processor other than the ALU. Fig. 7A and / or Figure 7B In at least one embodiment, the weight parameters may be stored in on-chip or off-chip memory and / or registers (shown or not shown) that configure the ALU of the graphics processor to execute one or more of the machine learning algorithms, neural network architectures, use cases, or training techniques described herein.

[0193] Such a component may be used to generate a single consistent tokenized description of at least a portion of a physical environment based in part on the set of observations and the aligned map data.

[0194] Fig.12 1 is a block diagram of a processor 1200 having one or more processor cores 1202A-1202N, an integrated memory controller 1214, and an integrated graphics processor 1208, according to at least one embodiment. In at least one embodiment, the processor 1200 may include additional cores, up to and including the additional core 1202N represented by the dashed box. In at least one embodiment, each processor core 1202A-1202N includes one or more internal cache units 1204A-1204N. In at least one embodiment, each processor core may also have access to one or more shared cache units 1206.

[0195] In at least one embodiment, the internal cache units 1204A-1204N and the shared cache unit 1206 represent a cache memory hierarchy within the processor 1200. In at least one embodiment, the cache memory units 1204A-1204N may include at least one level of instruction and data cache within each processor core and one or more levels of cache in a shared mid-level cache, such as level 2 (L2), level 3 (L3), level 4 (L4), or other levels of cache, where the highest level of cache before external memory is categorized as LLC. In at least one embodiment, cache alignment logic maintains alignment between the various cache units 1206 and 1204A-1204N.

[0196] In at least one embodiment, the processor 1200 may also include a set of one or more bus controller units 1216 and a system agent core 1210. In at least one embodiment, the one or more bus controller units 1216 manage a set of peripheral buses, such as one or more PCI or PCIe buses. In at least one embodiment, the system agent core 1210 provides management functions for various processor components. In at least one embodiment, the system agent core 1210 includes one or more integrated memory controllers 1214 to manage access to various external memory devices (not shown).

[0197] In at least one embodiment, one or more processor cores 1202A-1202N include support for multiple threads simultaneously. In at least one embodiment, system agent core 1210 includes components for coordinating processor cores 1202A-1202N during multithreaded processing. In at least one embodiment, system agent core 1210 may additionally include a power control unit (PCU) that includes logic and components for regulating one or more power states of processor cores 1202A-1202N and graphics processor 1208.

[0198] In at least one embodiment, the processor 1200 also includes a graphics processor 1208 for performing graphics processing operations. In at least one embodiment, the graphics processor 1208 is coupled to a shared cache unit 1206 and a system agent core 1210 including one or more integrated memory controllers 1214. In at least one embodiment, the system agent core 1210 also includes a display controller 1211 for driving the graphics processor output to one or more coupled displays. In at least one embodiment, the display controller 1211 may also be a separate module coupled to the graphics processor 1208 via at least one interconnect, or may be integrated within the graphics processor 1208.

[0199] In at least one embodiment, a ring-based interconnect unit 1212 is used to couple the internal components of the processor 1200. In at least one embodiment, alternative interconnect units may be used, such as point-to-point interconnects, switched interconnects, or other technologies. In at least one embodiment, the graphics processor 1208 is coupled to the ring-based interconnect unit 1212 via an I / O link 1213.

[0200] In at least one embodiment, I / O link 1213 represents at least one of a variety of I / O interconnects, including packaged I / O interconnects that facilitate communication between various processor components and high-performance embedded memory modules 1218 (e.g., eDRAM modules). In at least one embodiment, each of processor cores 1202A-1202N and graphics processor 1208 uses embedded memory modules 1218 as a shared last level cache.

[0201] In at least one embodiment, the processor cores 1202A-1202N are homogeneous cores that execute a common instruction set architecture. In at least one embodiment, the processor cores 1202A-1202N are heterogeneous in terms of instruction set architecture (ISA), wherein one or more processor cores 1202A-1202N execute a common instruction set, while one or more other processor cores 1202A-1202N execute a subset or a different instruction set of the common instruction set. In at least one embodiment, the processor cores 1202A-1202N are heterogeneous in terms of microarchitecture, wherein one or more cores with relatively high power consumption are coupled with one or more power cores with lower power consumption. In at least one embodiment, the processor 1200 can be implemented on one or more chips or implemented as a SoC integrated circuit.

[0202] Reasoning and / or training logic 715 is used to perform reasoning and / or training operations associated with one or more embodiments. Fig. 7A and / or Figure 7B 1200. In at least one embodiment, some or all of the inference and / or training logic 715 may be incorporated into the processor 1200. For example, in at least one embodiment, the training and / or inference techniques described herein may use one or more ALUs embodied in the graphics processor 1208, graphics cores 1202A-1202N, or Fig.12 In addition, in at least one embodiment, the reasoning and / or training operations described herein may use Fig. 7A and / or Figure 7B In at least one embodiment, the weight parameters may be stored in on-chip or off-chip memory and / or registers (shown or not shown) that configure the ALU of the graphics processor 1200 to execute one or more of the machine learning algorithms, neural network architectures, use cases, or training techniques described herein.

[0203] Such a component may be used to generate a single consistent tokenized description of at least a portion of a physical environment based in part on a set of observations and aligned map data.

[0204] Virtualized computing platform

[0205] Fig.13 is an example data flow diagram of a process 1300 for generating and deploying an image processing and inference pipeline according to at least one embodiment. In at least one embodiment, the process 1300 may be deployed for use with imaging devices, processing devices, and / or other device types at one or more facilities 1302. The process 1300 may be executed within a training system 1304 and / or a deployment system 1306. In at least one embodiment, the training system 1304 may be used to perform training, deployment, and implementation of machine learning models (e.g., neural networks, object detection algorithms, computer vision algorithms, etc.) for use in the deployment system 1306. In at least one embodiment, the deployment system 1306 may be configured to offload processing and computing resources between distributed computing environments to reduce infrastructure requirements at the facility 1302. In at least one embodiment, one or more applications in the pipeline may use or call services (e.g., inference, visualization, computation, AI, etc.) of the deployment system 1306 during execution of the application.

[0206] In at least one embodiment, some applications used in high-level processing and reasoning pipelines may use machine learning models or other AI to perform one or more processing steps. In at least one embodiment, the machine learning model can be trained at the facility 1302 using data 1308 (such as imaging data) generated at the facility 1302 (and stored on one or more picture archiving and communication system (PACS) servers at the facility 1302), can be trained using imaging or sequencing data 1308 from another (one or more) facilities, or a combination thereof. In at least one embodiment, the training system 1304 can be used to provide applications, services, and / or other resources for generating working, deployable machine learning models for the deployment system 1306.

[0207] In at least one embodiment, the model registry 1324 may be supported by an object store that may support versioning and object metadata. In at least one embodiment, the object store may be accessed from within the cloud platform through, for example, a cloud storage compatible application programming interface (API). In at least one embodiment, the machine learning models within the model registry 1324 may be uploaded, listed, modified, or deleted by developers or partners of the system that interface with the API. In at least one embodiment, the API may provide access to a method that allows a user to associate a model with an application using appropriate credentials so that the model may be executed as part of the execution of a containerized instantiation of the application.

[0208] In at least one embodiment, training pipeline 1304 ( Fig.13) may include scenarios where the facility 1302 is training its own machine learning model, or has an existing machine learning model that needs to be optimized or updated. In at least one embodiment, imaging data 1308 generated by one or more imaging devices, sequencing devices, and / or other device types may be received. In at least one embodiment, once the imaging data 1308 is received, the AI-assisted annotation 1310 may be used to help generate annotations corresponding to the imaging data 1308 to be used as ground truth data for the machine learning model. In at least one embodiment, the AI-assisted annotation 1310 may include one or more machine learning models (e.g., convolutional neural networks (CNNs)), which may be trained to generate annotations corresponding to certain types of imaging data 1308 (e.g., from certain devices). In at least one embodiment, the AI-assisted annotation 1310 may then be used directly, or may be adjusted or fine-tuned using an annotation tool to generate ground truth data. In at least one embodiment, the AI-assisted annotation 1310, the labeled data 1312, or a combination thereof may be used as ground truth data for training a machine learning model. In at least one embodiment, the trained machine learning model may be referred to as an output model 1316 and may be used by the deployment system 1306 as described herein.

[0209] In at least one embodiment, the training pipeline may include a scenario where the facility 1302 needs a machine learning model for performing one or more processing tasks of one or more applications in the deployment system 1306, but the facility 1302 may not currently have such a machine learning model (or may not have an optimized, efficient, or effective model for such a purpose). In at least one embodiment, an existing machine learning model may be selected from the model registry 1324. In at least one embodiment, the model registry 1324 may include machine learning models trained to perform a variety of different reasoning tasks on imaging data. In at least one embodiment, the machine learning models in the model registry 1324 may have been trained on imaging data from a facility different from the facility 1302 (e.g., a remotely located facility). In at least one embodiment, the machine learning model may have been trained on imaging data from one location, two locations, or any number of locations. In at least one embodiment, when training on imaging data from a particular location, the training may occur at that location, or at least in a manner that protects the confidentiality of the imaging data or restricts the transmission of the imaging data off-site. In at least one embodiment, once the model is trained or partially trained at one location, the machine learning model may be added to the model registry 1324. In at least one embodiment, the machine learning model can then be retrained or updated at any number of other facilities, and the retrained or updated model can be made available in model registry 1324. In at least one embodiment, the machine learning model can then be selected from model registry 1324 - and referred to as output model 1316 - and can be used in deployment system 1306 to perform one or more processing tasks for one or more applications of the deployment system.

[0210] In at least one embodiment, the scenario may include a facility 1302 that requires a machine learning model for performing one or more processing tasks for one or more applications in the deployment system 1306, but the facility 1302 may not currently have such a machine learning model (or may not have an optimized, efficient, or effective model for such a purpose). In at least one embodiment, the machine learning model selected from the model registry 1324 may not be fine-tuned or optimized for the imaging data 1308 generated at the facility 1302 due to differences in the population, the robustness of the training data used to train the machine learning model, the unusual diversity of the training data, and / or other issues with the training data. In at least one embodiment, AI-assisted annotations 1310 may be used to help generate annotations corresponding to the imaging data 1308, which are used as ground truth data for retraining or updating the machine learning model. In at least one embodiment, labeled data 1312 may be used as ground truth data for training the machine learning model. In at least one embodiment, retraining or updating the machine learning model may be referred to as model training 1314. In at least one embodiment, model training 1314 (e.g., AI-assisted annotations 1310, labeled data 1312, or a combination thereof) can be used as ground truth data for retraining or updating a machine learning model. In at least one embodiment, the trained machine learning model can be referred to as an output model 1316 and can be used by a deployment system 1306, as described herein.

[0211] In at least one embodiment, the deployment system 1306 may include software 1318, services 1320, hardware 1322, and / or other components, features, and functions. In at least one embodiment, the deployment system 1306 may include a software "stack" such that the software 1318 may be built on top of the services 1320 and may use the services 1320 to perform some or all of the processing tasks, and the services 1320 and software 1318 may be built on top of the hardware 1322 and use the hardware 1322 to perform the processing, storage, and / or other computing tasks of the deployment system 1306. In at least one embodiment, the software 1318 may include any number of different containers, each of which may perform an instantiation of an application. In at least one embodiment, each application may perform one or more processing tasks (e.g., reasoning, object detection, feature detection, segmentation, image enhancement, calibration, etc.) in a high-level processing and reasoning pipeline. In at least one embodiment, in addition to receiving and configuring imaging data for use by each container and / or containers used by facility 1302 after processing through the pipeline, a high-level processing and reasoning pipeline can also be defined based on the selection of different containers desired or required to process imaging data 1308 (e.g., to convert the output back into a usable data type. In at least one embodiment, the combination of containers within software 1318 (e.g., which constitute a pipeline) can be referred to as a virtual instrument (as described in more detail herein), and the virtual instrument can utilize services 1320 and hardware 1322 to perform some or all of the processing tasks of an application instantiated in a container.

[0212] In at least one embodiment, the data processing pipeline can receive input data (e.g., imaging data 1308) in a particular format in response to an inference request (e.g., a request from a user of the deployment system 1306). In at least one embodiment, the input data can represent one or more images, videos, and / or other data representations generated by one or more imaging devices. In at least one embodiment, the data can be pre-processed as part of the data processing pipeline to prepare the data for processing by one or more applications. In at least one embodiment, post-processing can be performed on the output of one or more inference tasks or other processing tasks of the pipeline to prepare output data for the next application and / or prepare the output data for transmission and / or use by the user (e.g., as a response to an inference request). In at least one embodiment, the inference task can be performed by one or more machine learning models, such as a trained or deployed neural network, which can include the output model 1316 of the training system 1304.

[0213] In at least one embodiment, tasks of a data processing pipeline may be encapsulated in containers, each container representing a discrete, fully functional instantiation of an application and a virtualized computing environment capable of referencing a machine learning model. In at least one embodiment, containers or applications may be published to a private (e.g., limited access) area of ​​a container registry (described in more detail herein), and trained or deployed models may be stored in the model registry 1324 and associated with one or more applications. In at least one embodiment, an image of an application (e.g., a container image) may be available in a container registry, and once a user selects an image from a container registry for deployment in a pipeline, the image may be used to generate an instantiated container for the application for use by the user's system.

[0214] In at least one embodiment, a developer (e.g., a software developer, a clinician, a physician, etc.) can develop, publish, and store applications (e.g., as containers) for performing image processing and / or reasoning on provided data. In at least one embodiment, the development, publishing, and / or storage can be performed using a software development kit (SDK) associated with the system (e.g., to ensure that the developed applications and / or containers conform to or are compatible with the system). In at least one embodiment, the developed applications can be tested locally (e.g., at a first facility, on data from the first facility) using the SDK, which is installed as part of the system (e.g., Fig.12 The processor 1200 in the container registry may support at least some of the services 1320. In at least one embodiment, because a DICOM object may contain from one to hundreds of images or other data types, and because the data varies, the developer may be responsible for managing (e.g., setting up constructs for building pre-processing into the application, etc.) the extraction and preparation of the incoming data. In at least one embodiment, once validated by the processor 1300 (e.g., for accuracy), the application is made available in the container registry for selection and / or implementation by the user to perform one or more processing tasks on the data at the user's facility (e.g., a second facility).

[0215] In at least one embodiment, the developer can then share the application or container over a network for use by a system (e.g., Fig.131300). In at least one embodiment, the completed and validated application or container can be stored in the container registry, and the associated machine learning model can be stored in the model registry 1324. In at least one embodiment, the requesting entity (which provides an inference or image processing request) can browse the container registry and / or the model registry 1324 to obtain applications, containers, data sets, machine learning models, etc., select the desired combination of elements to be included in the data processing pipeline, and submit the image processing request. In at least one embodiment, the request may include input data necessary to execute the request (and patient-related data in some examples), and / or may include a selection of applications and / or machine learning models to be executed when processing the request. In at least one embodiment, the request can then be passed to one or more components (e.g., a cloud) of the deployment system 1306 to perform processing of the data processing pipeline. In at least one embodiment, the processing performed by the deployment system 1306 may include referencing elements (e.g., applications, containers, models, etc.) selected from the container registry and / or the model registry 1324. In at least one embodiment, once the results are generated by the pipeline, the results may be returned to the user for reference (eg, for viewing in a viewing application suite executing locally, on a local workstation or terminal).

[0216] In at least one embodiment, in order to help process or execute applications or containers in the pipeline, services 1320 can be utilized. In at least one embodiment, services 1320 may include computing services, artificial intelligence (AI) services, visualization services, and / or other service types. In at least one embodiment, services 1320 may provide functions common to one or more applications in software 1318, so the functions may be abstracted into services that can be called or utilized by applications. In at least one embodiment, the functions provided by services 1320 may run dynamically and more efficiently, while also being well scaled by allowing applications to process data in parallel (e.g., using parallel computing platforms). In at least one embodiment, rather than requiring each application that shares the same functions provided by services 1320 to have a corresponding instance of services 1320, services 1320 may be shared between and among various applications. In at least one embodiment, as a non-limiting example, services 1320 may include an inference server or engine that can be used to perform detection or segmentation tasks. In at least one embodiment, a model training service may be included that may provide machine learning model training and / or retraining capabilities. In at least one embodiment, data enhancement services may be further included that may provide GPU accelerated data (e.g., DICOM, RIS, CIS, REST, RPC, raw, etc.) extraction, resizing, scaling, and / or other enhancements. In at least one embodiment, visualization services may be used that may add image rendering effects (e.g., ray tracing, rasterization, denoising, sharpening, etc.) to add realism to two-dimensional (2D) and / or three-dimensional (3D) models. In at least one embodiment, virtual instrument services may be included that provide beamforming, segmentation, reasoning, imaging, and / or support for other applications within the pipeline of the virtual instrument.

[0217] In at least one embodiment, where the service 1320 includes an AI service (e.g., an inference service), as part of the execution of an application, one or more machine learning models can be executed by calling (e.g., as an API call) an inference service (e.g., an inference server) to execute one or more machine learning models or processing thereof. In at least one embodiment, where another application includes one or more machine learning models for a segmentation task, the application can call the inference service to execute the machine learning model for performing one or more processing operations associated with the segmentation task. In at least one embodiment, the software 1318 implementing the high-level processing and inference pipeline, which includes a segmentation application and anomaly detection application, can be pipelined in that each application can call the same inference service to perform one or more inference tasks.

[0218] In at least one embodiment, the hardware 1322 may include a GPU, a CPU, a graphics card, an AI / deep learning system (e.g., an AI supercomputer such as NVIDIA's DGX), a cloud platform, or a combination thereof. In at least one embodiment, different types of hardware 1322 may be used to provide efficient, purpose-built support for the software 1318 and services 1320 in the deployment system 1306. In at least one embodiment, the use of GPU processing may be implemented to perform local processing (e.g., at the facility 1302) within the AI / deep learning system, in the cloud system, and / or other processing components of the deployment system 1306 to improve the efficiency, accuracy, and effectiveness of image processing and generation. In at least one embodiment, as a non-limiting example, with respect to deep learning, machine learning, and / or high-performance computing, the software 1318 and / or services 1320 may be optimized for GPU processing. In at least one embodiment, at least some of the computing environments of the deployment system 1306 and / or training system 1304 may be executed in a data center, one or more supercomputers, or high-performance computer systems with GPU-optimized software (e.g., a combination of hardware and software of an NVIDIA DGX system). In at least one embodiment, as described herein, hardware 1322 may include any number of GPUs that may be called to perform data processing in parallel. In at least one embodiment, the cloud platform may also include GPU optimized execution for deep learning tasks, GPU processing for machine learning tasks or other computing tasks. In at least one embodiment, an AI / deep learning supercomputer and / or GPU optimized software (e.g., as provided on NVIDIA's DGX system) may be used as a hardware abstraction and scaling platform to execute a cloud platform (e.g., NVIDIA's NGC). In at least one embodiment, the cloud platform may integrate an application container cluster system or coordination system (e.g., Kubernetes) on multiple GPUs to allow seamless scaling and load balancing.

[0219] Fig.14 is a system diagram of an example system 1400 for generating and deploying an imaging deployment pipeline according to at least one embodiment. In at least one embodiment, the system 1400 can be used to implement Fig.13 The process 1300 and / or other processes of the system 1400 may include a high-level processing and reasoning pipeline. In at least one embodiment, the system 1400 may include a training system 1304 and a deployment system 1306. In at least one embodiment, the training system 1304 and the deployment system 1306 may be implemented using software 1318, services 1320, and / or hardware 1322, as described herein.

[0220] In at least one embodiment, system 1400 (e.g., training system 1304 and / or deployment system 1306) can be implemented in a cloud computing environment (e.g., using cloud 1426). In at least one embodiment, system 1400 can be implemented locally (with respect to a healthcare service facility) or as a combination of cloud computing resources and local computing resources. In at least one embodiment, access to APIs in cloud 1426 can be restricted to authorized users by establishing security measures or protocols. In at least one embodiment, the security protocol can include a network token, which can be signed by an authentication (e.g., AuthN, AuthZ, Gluecon, etc.) service and can carry appropriate authorization. In at least one embodiment, the API of the virtual instrument (described herein) or other instances of system 1400 can be restricted to a set of public IPs that have been audited or authorized for interaction.

[0221] In at least one embodiment, the various components of system 1400 can communicate with each other using any of a variety of different network types, including but not limited to local area networks (LANs) and / or wide area networks (WANs) via wired and / or wireless communication protocols. In at least one embodiment, communications between facilities and components of system 1400 (e.g., for sending inference requests, for receiving results of inference requests, etc.) can be transmitted via one or more data buses, wireless data protocols (Wi-Fi), wired data protocols (e.g., Ethernet), etc.

[0222] In at least one embodiment, similar to the present disclosure regarding Fig.13 As described, the training system 1304 can execute the training pipeline 1404. In at least one embodiment, where the deployment system 1306 will use one or more machine learning models in the deployment pipeline 1410, the training pipeline 1404 can be used to train or retrain one or more (e.g., pre-trained) models, and / or implement one or more pre-trained models 1406 (e.g., without retraining or updating). In at least one embodiment, as a result of the training pipeline 1404, an output model 1316 can be generated. In at least one embodiment, the training pipeline 1404 can include any number of processing steps, such as, but not limited to, conversion or adaptation of imaging data (or other input data). In at least one embodiment, different training pipelines 1404 can be used for different machine learning models used by the deployment system 1306. In at least one embodiment, similar to the description regarding Fig.13 The first example training pipeline 1404 described may be used for a first machine learning model, similar to the training pipeline 1404 described with respect to Fig.13 The second example training pipeline 1404 described may be used for a second machine learning model, similar to the one described with respect to Fig.13The training pipeline 1404 of the third example described may be used for a third machine learning model. In at least one embodiment, any combination of tasks within the training system 1304 may be used according to the requirements of each respective machine learning model. In at least one embodiment, one or more machine learning models may have been trained and are ready for deployment, so the training system 1304 may not perform any processing on the machine learning model, and one or more machine learning models may be implemented by the deployment system 1306.

[0223] In at least one embodiment, the output model 1316 and / or the pre-trained model 1406 may include any type of machine learning model, depending on the implementation or embodiment. In at least one embodiment and without limitation, the machine learning model used by the system 1400 may include using linear regression, logistic regression, decision trees, support vector machines (SVM), naive Bayes, k-nearest neighbors (Knn), k-means clustering, random forests, dimensionality reduction algorithms, gradient boosting algorithms, neural networks (e.g., autoencoders, convolutions, recursion, perceptrons, long / short term memory (LSTM), Hopfield, Boltzmann, deep belief, deconvolution, generative adversarial, liquid state machines, etc.), and / or other types of machine learning models.

[0224] In at least one embodiment, the training pipeline 1404 may include AI-assisted annotation, as described herein with respect to at least Fig.14In at least one embodiment, the labeled data 1312 may be generated by any number of techniques (e.g., traditional annotations). In at least one embodiment, in some examples, labels or other annotations may be generated in a drawing program (e.g., an annotation program), a computer-aided design (CAD) program, a labeling program, another type of application suitable for generating annotations or labels for ground truth, and / or may be hand-drawn. In at least one embodiment, ground truth data may be synthetically generated (e.g., generated from a computer model or rendering), realistically generated (e.g., designed and generated from real-world data), machine-generated (e.g., extracting features from data using feature analysis and learning and then generating labels), manually annotated (e.g., a labeler or annotation expert, defining the location of the label), and / or a combination thereof. In at least one embodiment, for each instance of imaging data 1308 (or other data types used by machine learning models), there may be corresponding ground truth data generated by the training system 1304. In at least one embodiment, AI-assisted annotation 1310 may be performed as part of a deployment pipeline 1410; supplementing or replacing AI-assisted annotation 1310 included in the training pipeline 1404. In at least one embodiment, the system 1400 may include a multi-layer platform that may include a software layer (e.g., software 1318) of a diagnostic application (or other application type) that may perform one or more medical imaging and diagnostic functions. In at least one embodiment, the system 1400 may be communicatively coupled (e.g., via an encrypted link) to a PACS server network of one or more facilities. In at least one embodiment, the system 1400 may be configured to access and reference data from a PACS server to perform operations such as training a machine learning model, deploying a machine learning model, image processing, reasoning, and / or other operations.

[0225] In at least one embodiment, the software layer may be implemented as a secure, encrypted, and / or authenticated API through which an application or container may be invoked (e.g., called) from an external environment (e.g., facility 1302). In at least one embodiment, the application may then call or execute one or more services 1320 to perform computational, AI, or visualization tasks associated with the respective application, and the software 1318 and / or services 1320 may utilize hardware 1322 to perform processing tasks in an effective and efficient manner. In at least one embodiment, a pair of DICOM adapters 1402A, 1402B may be used for communications sent to or received by the training system 1304 and the deployment system 1306.

[0226] In at least one embodiment, the deployment system 1306 can execute a deployment pipeline 1410. In at least one embodiment, the deployment pipeline 1410 can include any number of applications, which can be sequential, non-sequential, or otherwise applied to imaging data (and / or other data types) - including AI-assisted annotations, generated by imaging devices, sequencing devices, genomics devices, etc., as described above. In at least one embodiment, as described herein, the deployment pipeline 1410 for an individual device can be referred to as a virtual instrument for the device (e.g., a virtual ultrasound instrument, a virtual CT scanning instrument, a virtual sequencing instrument, etc.). In at least one embodiment, there can be more than one deployment pipeline 1410 for a single device, depending on the information desired from the data generated by the device. In at least one embodiment, there can be a first deployment pipeline 1410 in the case where an abnormality is desired to be detected from an MRI machine, and there can be a second deployment pipeline 1410 in the case where image enhancement is desired from the output of the MRI machine.

[0227] In at least one embodiment, the image generation application may include a processing task that includes the use of a machine learning model. In at least one embodiment, a user may wish to use their own machine learning model, or select a machine learning model from the model registry 1324. In at least one embodiment, a user may implement their own machine learning model or select a machine learning model to be included in an application that performs a processing task. In at least one embodiment, the application may be selectable and customizable, and by defining the construction of the application, the deployment and implementation of the application for a particular user is presented as a more seamless user experience. In at least one embodiment, by leveraging other features of the system 1400 (e.g., services 1320 and hardware 1322), the deployment pipeline 1410 can be more user-friendly, provide easier integration, and produce more accurate, efficient, and timely results.

[0228] In at least one embodiment, deployment system 1306 may include a user interface ("UI") 1414 (e.g., a graphical user interface, a web interface, etc.) that may be used to select applications to be included in deployment pipeline 1410, to arrange applications, to modify or change applications or their parameters or configuration, to use and interact with deployment pipeline 1410 during setup and / or deployment, and / or to otherwise interact with deployment system 1306. In at least one embodiment, although not shown with respect to training system 1304, UI 1414 (or a different user interface) may be used to select models to use in deployment system 1306, to select models for training or retraining in training system 1304, and / or to otherwise interact with training system 1304.

[0229] In at least one embodiment, in addition to the application coordination system 1428, a pipeline manager 1412 may be used to manage the interaction between the application or container of the deployment pipeline 1410 and the service 1320 and / or the hardware 1322. In at least one embodiment, the pipeline manager 1412 may be configured to facilitate the interaction from application to application, from application to service 1320, and / or from application or service to hardware 1322. In at least one embodiment, although shown as included in the software 1318, this is not intended to be limiting, and in some examples, the pipeline manager 1412 may be included in the service 1320. In at least one embodiment, the application coordination system 1428 (e.g., Kubernetes, DOCKER, etc.) may include a container coordination system that can group applications into containers as a logical unit for coordination, management, scaling, and deployment. In at least one embodiment, by associating applications from the deployment pipeline 1410 (e.g., rebuilding applications, splitting applications, etc.) with individual containers, each application can be executed in a self-contained environment (e.g., at the kernel level) to improve speed and efficiency.

[0230] In at least one embodiment, each application and / or container (or its image) can be developed, modified, and deployed separately (e.g., a first user or developer can develop, modify, and deploy a first application, and a second user or developer can develop, modify, and deploy a second application separate from the first user or developer), which can allow for focus and attention on the tasks of a single application and / or container without being hindered by the tasks of another application or container. In at least one embodiment, the pipeline manager 1412 and the application coordination system 1428 can assist in communication and collaboration between different containers or applications. In at least one embodiment, as long as the expected inputs and / or outputs of each container or application are known to the system (e.g., based on the construction of the application or container), the application coordination system 1428 and / or the pipeline manager 1412 can facilitate communication and sharing of resources between and among each application or container. In at least one embodiment, since one or more applications or containers in the deployment pipeline 1410 can share the same services and resources, the application coordination system 1428 can coordinate, load balance, and determine the sharing of services or resources between and among the various applications or containers. In at least one embodiment, the scheduler can be used to track resource requirements of applications or containers, current or planned use of those resources, and resource availability. Thus, in at least one embodiment, the scheduler can allocate resources to different applications and allocate resources between and among applications, taking into account the needs and availability of the system. In some examples, the scheduler (and / or other components of the application coordination system 1428) can determine resource availability and distribution based on constraints imposed on the system (e.g., user constraints), such as quality of service (QoS), the urgency of data output (e.g., to determine whether to perform real-time processing or delayed processing), etc.

[0231] In at least one embodiment, the services 1320 utilized and shared by the applications or containers in the deployment system 1306 may include computing services 1416, AI services 1418, visualization services 1420, and / or other service types. In at least one embodiment, an application may call (e.g., execute) one or more services 1320 to perform processing operations for the application. In at least one embodiment, an application may utilize computing services 1416 to perform supercomputing or other high performance computing (HPC) tasks. In at least one embodiment, one or more computing services 1416 may be utilized to perform parallel processing (e.g., using a parallel computing platform 1430) to process data substantially simultaneously by one or more applications and / or one or more tasks of a single application. In at least one embodiment, a parallel computing platform 1430 (e.g., NVIDIA's CUDA) may allow general purpose computing (GPGPU) on a GPU (e.g., GPU / Graphic 1422). In at least one embodiment, the software layer of the parallel computing platform 1430 may provide access to a virtual instruction set and parallel computing elements of a GPU to execute computing kernels. In at least one embodiment, the parallel computing platform 1430 may include memory, and in some embodiments, memory may be shared between and among multiple containers, and / or between and among different processing tasks within a single container. In at least one embodiment, inter-process communication (IPC) calls may be generated for multiple containers and / or multiple processes within a container to use the same data from a shared memory segment of the parallel computing platform 1430 (e.g., where multiple different stages of an application or multiple applications are processing the same information). In at least one embodiment, instead of copying data and moving the data to different locations in memory (e.g., read / write operations), the same data in the same location of the memory may be used for any number of processing tasks (e.g., at the same time, at different times, etc.). In at least one embodiment, since the data as a result of the processing is used to generate new data, this information of the new location of the data may be stored and shared between various applications. In at least one embodiment, the location of the data and the location of the updated or modified data may be part of the definition of how to understand the payload in the container.

[0232] In at least one embodiment, the AI ​​service 1418 may be utilized to perform an inference service for executing a machine learning model associated with an application (e.g., a task is to execute one or more processing tasks of an application). In at least one embodiment, the AI ​​service 1418 may utilize the AI ​​system 1424 to execute a machine learning model (e.g., a neural network such as a CNN) for segmentation, reconstruction, object detection, feature detection, classification, and / or other inference tasks. In at least one embodiment, the application of the deployment pipeline 1410 may use one or more output models 1316 from the training system 1304 and / or other models of the application to perform inference on the imaging data. In at least one embodiment, two or more examples of inference using the application coordination system 1428 (e.g., a scheduler) may be available. In at least one embodiment, the first category may include a high priority / low latency path that can achieve a higher service level agreement, such as for performing inference on an urgent request in an emergency, or for a radiologist during a diagnostic process. In at least one embodiment, the second category may include a standard priority path that may be used for requests that may not be urgent or where analysis can be performed at a later time. In at least one embodiment, application coordination system 1428 can allocate resources (e.g., services 1320 and / or hardware 1322) for different reasoning tasks of AI service 1418 based on priority paths.

[0233] In at least one embodiment, the shared memory may be installed to the AI ​​service 1418 in the system 1400. In at least one embodiment, the shared memory may operate as a cache (or other storage device type) and may be used to process inference requests from applications. In at least one embodiment, when an inference request is submitted, a set of API instances of the deployment system 1306 may receive the request and may select one or more instances (e.g., for best fit, for load balancing, etc.) to process the request. In at least one embodiment, in order to process the request, the request may be entered into a database, and if it is not already in the cache, the machine learning model may be located from the model registry 1324, a verification step may ensure that the appropriate machine learning model is loaded into the cache (e.g., shared storage), and / or a copy of the model may be saved to the cache. In at least one embodiment, if the application is not yet running or there are not enough instances of the application, a scheduler (e.g., a scheduler of the pipeline manager 1412) may be used to start the application referenced in the request. In at least one embodiment, if the inference server has not yet been started to execute the model, the inference server may be started. Any number of inference servers may be started for each model. In at least one embodiment, in a pull model that clusters inference servers, the model can be cached whenever load balancing is beneficial. In at least one embodiment, the inference servers can be statically loaded into the corresponding distributed servers.

[0234] In at least one embodiment, inference can be performed using an inference server running in a container. In at least one embodiment, an instance of an inference server can be associated with a model (and optionally with multiple versions of a model). In at least one embodiment, if an instance of an inference server does not exist when a request is received to perform inference on a model, a new instance can be loaded. In at least one embodiment, when the inference server is started, the model can be passed to the inference server so that the same container can be used to serve different models as long as the inference server is running as a different instance.

[0235] In at least one embodiment, during application execution, a request for reasoning for a given application may be received, and a container (e.g., an instance of a hosting reasoning server) may be loaded (if not already loaded), and a launcher may be called. In at least one embodiment, preprocessing logic in the container may load, decode, and / or perform any additional preprocessing on the incoming data (e.g., using a CPU and / or GPU). In at least one embodiment, once the data is ready for reasoning, the container may perform reasoning on the data as needed. In at least one embodiment, this may include a single reasoning call for one image (e.g., a hand X-ray), or may require reasoning for hundreds of images (e.g., a chest CT). In at least one embodiment, the application may summarize the results before completion, which may include, but is not limited to, a single confidence score, pixel-level segmentation, voxel-level segmentation, generating visualizations, or generating text to summarize the results. In at least one embodiment, different priorities may be assigned to different models or applications. For example, some models may have a real-time (TAT less than 1 minute) priority, while other models may have a lower priority (e.g., TAT less than 10 minutes). In at least one embodiment, model execution time may be measured from a requesting mechanism or entity and may include collaborative network traversal time as well as execution time of an inference service.

[0236] In at least one embodiment, the transmission of requests between the service 1320 and the reasoning application can be hidden behind a software development kit (SDK), and robust transmission can be provided through queues. In at least one embodiment, requests will be placed in queues through an API for individual application / tenant ID combinations, and the SDK will pull requests from the queue and provide the request to the application. In at least one embodiment, the name of the queue can be provided in the environment from which the SDK will pick up the queue. In at least one embodiment, asynchronous communication through queues may be useful because it can allow any instance of the application to pick up work when it is available. Results can be transmitted back through the queue to ensure that no data is lost. In at least one embodiment, the queue can also provide the ability to split the work, because the highest priority work can enter the queue connected to most instances of the application, and the lowest priority work can enter the queue connected to a single instance, which processes the tasks in the order received. In at least one embodiment, the application can run on a GPU-accelerated instance generated in the cloud 1426, and the reasoning service can perform reasoning on the GPU.

[0237] In at least one embodiment, visualization services 1420 may be utilized to generate visualizations for viewing application and / or deployment pipeline 1410 outputs. In at least one embodiment, visualization services 1420 may utilize GPU / Graphics 1422 to generate visualizations. In at least one embodiment, visualization services 1420 may implement rendering effects such as ray tracing to generate higher quality visualizations. In at least one embodiment, visualizations may include, but are not limited to, 2D image rendering, 3D volume rendering, 3D volume reconstruction, 2D tomographic slices, virtual reality displays, augmented reality displays, and the like. In at least one embodiment, a virtualized environment may be used to generate a virtual interactive display or environment (e.g., a virtual environment) for system users (e.g., doctors, nurses, radiologists, etc.) to interact with. In at least one embodiment, visualization services 1420 may include internal visualizers, movies, and / or other rendering or image processing capabilities or functions (e.g., ray tracing, rasterization, internal optics, etc.).

[0238] In at least one embodiment, hardware 1322 may include GPU / Graphic 1422, AI system 1424, cloud 1426, and / or any other hardware for executing training system 1304 and / or deployment system 1306. In at least one embodiment, GPU / Graphic 1422 (e.g., NVIDIA's TESLA and / or QUADRO GPUs) may include any number of GPUs that may be used to perform processing tasks for any feature or functionality of compute service 1416, AI service 1418, visualization service 1420, other services, and / or software 1318. For example, for AI service 1418, GPU / Graphic 1422 may be used to perform pre-processing on imaging data (or other data types used by machine learning models), post-processing on the output of machine learning models, and / or perform inference (e.g., to execute machine learning models). In at least one embodiment, cloud 1426, AI system 1424, and / or other components of system 1400 may use GPU / Graphic 1422. In at least one embodiment, cloud 1426 may include a GPU-optimized platform for deep learning tasks. In at least one embodiment, AI system 1424 may use a GPU, and cloud 1426 (or at least part of a task being deep learning or reasoning) may be executed using one or more AI systems 1424. Likewise, while hardware 1322 is shown as discrete components, this is not intended to be limiting, and any component of hardware 1322 may be combined with or utilized by any other component of hardware 1322.

[0239] In at least one embodiment, AI system 1424 may include a purpose-built computing system (e.g., a supercomputer or HPC) configured for reasoning, deep learning, machine learning, and / or other artificial intelligence tasks. In at least one embodiment, in addition to CPU, RAM, storage, and / or other components, features, or functions, AI system 1424 (e.g., NVIDIA's DGX) may also include software (e.g., a software stack) that can use multiple GPUs / Graphics 1422 to perform GPU-optimized tasks. In at least one embodiment, one or more AI systems 1424 may be implemented in a cloud 1426 (e.g., in a data center) to perform some or all of the AI-based processing tasks of system 1400.

[0240] In at least one embodiment, cloud 1426 may include a GPU-accelerated infrastructure (e.g., NVIDIA's NGC) that may provide a GPU-optimized platform for executing processing tasks of system 1400. In at least one embodiment, cloud 1426 may include an AI system 1424 for executing one or more AI-based tasks of system 1400 (e.g., as a hardware abstraction and scaling platform). In at least one embodiment, cloud 1426 may be integrated with an application coordination system 1428 that utilizes multiple GPUs to allow seamless scaling and load balancing between and among applications and services 1320. In at least one embodiment, cloud 1426 may be responsible for executing at least some services 1320 of system 1400, including compute services 1416, AI services 1418, and / or visualization services 1420, as described herein. In at least one embodiment, cloud 1426 can perform large and small batch inference (e.g., executing NVIDIA's TENSOR RT), provide accelerated parallel computing APIs and platforms 1430 (e.g., NVIDIA's CUDA), execute application coordination system 1428 (e.g., KUBERNETES), provide graphics rendering APIs and platforms (e.g., for ray tracing, 2D graphics, 3D graphics, and / or other rendering techniques to produce higher quality movie effects), and / or can provide other functionality for system 1400.

[0241] Fig.15A A data flow diagram of a process 1500 for training, retraining, or updating a machine learning model according to at least one embodiment is shown. In at least one embodiment, a non-limiting example may be used. Fig.14The process 1500 may be performed by the system 1400 of FIG. 1400. In at least one embodiment, the process 1500 may utilize services and / or hardware as described herein. In at least one embodiment, the refined model 1512 generated by the process 1500 may be executed by the deployment system for one or more containerized applications in the deployment pipeline 1510.

[0242] In at least one embodiment, model training 1514 may include retraining or updating an initial model 1504 (e.g., a pre-trained model) using new training data (e.g., new input data (such as customer data set 1506), and / or new ground truth data associated with the input data). In at least one embodiment, to retrain or update the initial model 1504, the output or loss layer of the initial model 1504 may be reset, deleted, and / or replaced with an updated or new output or loss layer. In at least one embodiment, the initial model 1504 may have previously fine-tuned parameters (e.g., weights and / or biases) retained from previous training, so training or retraining 1514 may not take as long or require as much processing as training the model from scratch. In at least one embodiment, during model training, by resetting or replacing the output or loss layer of the initial model 1504, when generating predictions on the new customer data set 1506, the parameters for the new data set may be updated and re-adjusted based on the loss calculation associated with the accuracy of the output or loss layer.

[0243] In at least one embodiment, the pre-trained model 1506 may be stored in a data store or registry. In at least one embodiment, the pre-trained model 1506 may have been trained at least in part at one or more facilities other than the facility performing process 1500. In at least one embodiment, in order to protect the privacy and rights of patients, subjects, or customers of different facilities, the pre-trained model 1506 may have been trained locally using locally generated customer or patient data. In at least one embodiment, the cloud and / or other hardware may be used to train the pre-trained model 1506, but confidential, privacy-protected patient data may not be transmitted to, used by, or accessed by any component of the cloud (or other non-local hardware). In at least one embodiment, if the pre-trained model 1506 is trained using patient data from more than one facility, the pre-trained model 1506 may have been trained separately for each facility before training on patient or customer data from another facility. In at least one embodiment, customer or patient data from any number of facilities may be used to train pre-trained model 1506 on-site and / or off-site, such as in a data center or other cloud computing infrastructure, for example where the customer or patient data has been issued privacy issues (e.g., by being waived, used for experimental purposes, etc.), or where the customer or patient data is included in a public dataset.

[0244] In at least one embodiment, when selecting an application to use in a deployment pipeline, the user may also select a machine learning model to use for a particular application. In at least one embodiment, the user may not have a model to use, so the user may select a pre-trained model 1506 to use with the application. In at least one embodiment, the pre-trained model 1506 may not be optimized to generate accurate results on the customer data set 1506 of the user facility (e.g., based on patient diversity, demographics, type of medical imaging equipment used, etc.). In at least one embodiment, the pre-trained model 1506 may be updated, retrained, and / or fine-tuned for use at various facilities before deploying the pre-trained model into a deployment pipeline for use with one or more applications.

[0245] In at least one embodiment, a user may select a pre-trained model 1506 to be updated, retrained, and / or fine-tuned, and the pre-trained model may be referred to as the initial model 1504 for training the system in process 1500. In at least one embodiment, a customer data set 1506 (e.g., imaging data, genomic data, sequencing data, or other data types generated by equipment at a facility) may be used to perform model training (which may include, but is not limited to, transfer learning) on ​​the initial model 1504 to generate a refined model 1512. In at least one embodiment, ground truth data corresponding to the customer data set 1506 may be generated by the model training system 1304. In at least one embodiment, the ground truth data may be generated at least in part by a clinician, scientist, physician, practitioner at the facility.

[0246] In at least one embodiment, AI-assisted annotation 1310 may be used in some examples to generate ground truth data. In at least one embodiment, AI-assisted annotation 1310 (e.g., implemented using an AI-assisted annotation SDK) may utilize a machine learning model (e.g., a neural network) to generate ground truth data for recommendations or predictions for a customer data set. In at least one embodiment, a user may use the annotation tool within a user interface (graphical user interface (GUI)) on a computing device.

[0247] In at least one embodiment, user 1510 can interact with the GUI via computing device 1508 to edit or fine-tune (automatic) annotations. In at least one embodiment, a polygon editing feature can be used to move vertices of a polygon to a more precise or fine-tuned position.

[0248] In at least one embodiment, once the customer data set 1506 has associated ground truth data, the ground truth data (e.g., from AI-assisted annotations, manual labeling, etc.) can be used during model training to generate a refined model 1512. In at least one embodiment, the customer data set 1506 can be applied to the initial model 1504 any number of times, and the ground truth data can be used to update the parameters of the initial model 1504 until an acceptable level of accuracy is achieved for the refined model 1512. In at least one embodiment, once the refined model 1512 is generated, the refined model 1512 can be deployed within one or more deployment pipelines at a facility for use in performing one or more processing tasks with respect to medical imaging data.

[0249] In at least one embodiment, the refined model 1512 can be uploaded to the pre-trained models in the model registry to be selected by another facility. In at least one embodiment, this process can be completed at any number of facilities, so that the refined model 1512 can be further refined any number of times on new data sets to generate a more general model.

[0250] Fig. 15B is an example illustration of a client-server architecture 1532 for enhancing an annotation tool with a pre-trained annotation model 1542 according to at least one embodiment. In at least one embodiment, an AI-assisted annotation tool 1536 can be instantiated based on the client-server architecture 1532. In at least one embodiment, the AI-assisted annotation tool 1536 in an imaging application can assist a radiologist, for example, in identifying organs and abnormalities. In at least one embodiment, the imaging application can include a software tool that, as a non-limiting example, helps a user 1510 identify several extreme points on a specific organ of interest in an original image 1534 (e.g., in a 3D MRI or CT scan) and receive automatic annotation results for all 2D slices of the specific organ. In at least one embodiment, the results can be stored in a data store as training data 1538 and used as, for example, but not limited to, ground truth data for training. In at least one embodiment, when a computing device 1508 sends extreme points for AI-assisted annotation, for example, a deep learning model can receive the data as input and return an inference result for segmenting an organ or anomaly. In at least one embodiment, a pre-instantiated annotation tool (e.g., Fig. 15B The AI-assisted annotation tool 1536 in the example may be enhanced by making an API call (e.g., API call 1544) to a server (such as an annotation assistant server 1540), which may include a set of pre-trained models 1542 stored in, for example, an annotation model registry. In at least one embodiment, the annotation model registry may store pre-trained models 1542 (e.g., machine learning models, such as deep learning models) that are pre-trained to perform AI-assisted annotation 1310 on specific organs or abnormalities. These models may be further updated using a training pipeline. In at least one embodiment, the pre-installed annotation tools may be improved over time as new labeled data is added.

[0251] Autonomous Vehicles

[0252] Fig.16A An example of an autonomous vehicle 1600 is shown in accordance with at least one embodiment. In at least one embodiment, autonomous vehicle 1600 (alternatively referred to herein as "vehicle 1600") may be, but is not limited to, a passenger vehicle, such as a car, truck, bus, and / or another type of vehicle that may accommodate one or more passengers. In at least one embodiment, vehicle 1600 may be a semi-tractor-trailer for hauling cargo. In at least one embodiment, vehicle 1600 may be an aircraft, a robotic vehicle, or another type of vehicle.

[0253] An autonomous vehicle may be described according to the automation levels defined by the National Highway Traffic Safety Administration (“NHTSA”) of the U.S. Department of Transportation and the Society of Automotive Engineers (“SAE”) in the “Taxonomy and Definitions for Terms Related to Driving Automation Systems for On-Road Motor Vehicles” (e.g., Standard No. J3016-201806, dated June 15, 2018, Standard No. J3016-201609, dated September 30, 2016, and previous and future versions of this standard). In at least one embodiment, the vehicle 1600 may be capable of functioning according to one or more of the levels 1 to 5 of the autonomous driving levels. For example, in at least one embodiment, depending on the embodiment, the vehicle 1600 may be capable of conditional automation (level 3), high automation (level 4), and / or full automation (level 5).

[0254] In at least one embodiment, the vehicle 1600 may include, but is not limited to, components such as a chassis, a body, wheels (e.g., 2, 4, 6, 8, 18, etc.), tires, axles, and other components of the vehicle. In at least one embodiment, the vehicle 1600 may include, but is not limited to, a propulsion system 1650, such as an internal combustion engine, a hybrid power plant, an all-electric engine, and / or another type of propulsion system. In at least one embodiment, the propulsion system 1650 may be connected to a drive train of the vehicle 1600, which may include, but is not limited to, a transmission, so that the vehicle 1600 can be propelled. In at least one embodiment, the propulsion system 1650 may be controlled in response to receiving a signal from a throttle / accelerator 1652.

[0255] In at least one embodiment, when propulsion system 1650 is operating (e.g., when vehicle 1600 is traveling), steering system 1654 (which may include, but is not limited to, a steering wheel) is used to steer vehicle 1600 (e.g., along a desired path or route). In at least one embodiment, steering system 1654 may receive signals from steering actuator 1656. In at least one embodiment, the steering wheel may be optional for full automation (Level 5) functionality. In at least one embodiment, brake sensor system 1646 may be used to operate vehicle brakes in response to signals received from brake actuator 1648 and / or brake sensors.

[0256] In at least one embodiment, controller 1636 may include, but is not limited to, one or more system on a chip (“SoC”) ( Fig.16A1600) and / or a graphics processing unit ("GPU") provide signals (e.g., representing commands) to one or more components and / or systems of the vehicle 1600. For example, in at least one embodiment, the controller 1636 can send signals to operate vehicle brakes through the brake actuator 1648, operate the steering system 1654 through one or more steering actuators 1656, and operate the propulsion system 1650 through one or more throttles / accelerators 1652. In at least one embodiment, the one or more controllers 1636 may include one or more onboard (e.g., integrated) computing devices that process sensor signals and output operating commands (e.g., signals representing commands) to achieve autonomous driving and / or assist the driver in driving the vehicle 1600. In at least one embodiment, the one or more controllers 1636 may include a first controller for autonomous driving functions, a second controller for functional safety functions, a third controller for artificial intelligence functions (e.g., computer vision), a fourth controller for infotainment functions, a fifth controller for redundancy in emergency situations, and / or other controllers. In at least one embodiment, a single controller may handle two or more of the above functions, two or more controllers may handle a single function, and / or any combination thereof.

[0257] In at least one embodiment, the one or more controllers 1636 provide signals for controlling one or more components and / or systems of the vehicle 1600 in response to sensor data received from one or more sensors (e.g., sensor inputs). In at least one embodiment, the sensor data can be received from sensors such as, but not limited to, one or more global navigation satellite system ("GNSS") sensors 1658 (e.g., one or more global positioning system sensors), one or more RADAR sensors 1660, one or more ultrasonic sensors 1662, one or more LIDAR sensors 1664, one or more inertial measurement unit (IMU) sensors 1666 (e.g., one or more accelerometers, one or more gyroscopes, one or more magnetic compasses, one or more magnetometers, etc.), one or more microphones 1696, one or more stereo cameras 1668, one or more wide angle cameras 1670 (e.g., fisheye cameras), one or more infrared cameras 1672, one or more surround cameras 1674 (e.g., 360 degree cameras), telephoto cameras (e.g., infrared cameras), and / or a plurality of other sensors. Fig.16A Not shown), mid-range camera ( Fig.16A1600), one or more speed sensors 1644 (e.g., for measuring the speed of the vehicle 1600), one or more vibration sensors 1642, one or more steering sensors 1640, one or more brake sensors (e.g., as part of a brake sensor system 1646), and / or other sensor types are received.

[0258] In at least one embodiment, one or more controllers 1636 may receive input (e.g., represented by input data) from a dashboard 1632 of the vehicle 1600 and provide output (e.g., represented by output data, display data, etc.) via a human machine interface (“HMI”) display 1634, an audible annunciator, a speaker, and / or other components of the vehicle 1600. In at least one embodiment, the output may include information such as vehicle speed, velocity, time, map data (e.g., high definition map ( Fig.16A 1600 ), location data (e.g., the location of the vehicle 1600 , such as on a map), directions, the locations of other vehicles (e.g., occupancy barriers), information about objects and the states of objects sensed by one or more controllers 1636 , etc. For example, in at least one embodiment, HMI display 1634 can display information about the presence of one or more objects (e.g., road signs, warning signs, traffic signal changes, etc.) and / or information about driving maneuvers the vehicle has, is, or will make (e.g., changing lanes now, taking exit 34B in two miles, etc.).

[0259] In at least one embodiment, the vehicle 1600 also includes a network interface 1624, which can use one or more wireless antennas 1626 and / or one or more modems to communicate through one or more networks. For example, in at least one embodiment, the network interface 1624 may be able to communicate through long-term evolution ("LTE"), wideband code division multiple access ("WCDMA"), universal mobile telecommunications system ("UMTS"), global system for mobile communications ("GSM"), IMT-CDMA multi-carrier ("CDMA2000") networks, etc. In at least one embodiment, one or more wireless antennas 1626 can also use one or more local area networks (e.g., Bluetooth, Bluetooth Low Energy (LE), Z-Wave, ZigBee, etc.) and / or one or more low power wide area networks (hereinafter referred to as "LPWAN") (e.g., LoRaWAN, SigFox, etc.) protocols to enable communication between objects in the environment (e.g., vehicles, mobile devices).

[0260] Reasoning and / or training logic 715 is used to perform reasoning and / or training operations associated with one or more embodiments. Fig. 7A and / or Figure 7B Provides details about the reasoning and / or training logic 715. In at least one embodiment, the reasoning and / or training logic 715 may be implemented in the system Fig.16A for inferring or predicting operations based at least in part on weight parameters computed using neural network training operations, neural network functions and / or architectures, or neural network use cases described herein.

[0261] Such a component may be used to generate a single consistent tokenized description of at least a portion of a physical environment based in part on a set of observations and aligned map data.

[0262] Fig. 16B According to at least one embodiment, Fig.16A An example of camera positions and fields of view for an autonomous vehicle 1600 is shown. In at least one embodiment, the cameras and respective fields of view are an example embodiment and are not intended to be limiting. For example, in at least one embodiment, additional and / or alternative cameras may be included and / or the cameras may be located at different locations on the vehicle 1600.

[0263] In at least one embodiment, the camera type for the camera may include, but is not limited to, a digital camera that may be suitable for use with components and / or systems of the vehicle 1600. In at least one embodiment, one or more cameras may operate at an automotive safety integrity level ("ASIL") B and / or other ASILs. In at least one embodiment, the camera type may have any image capture rate, such as 60 frames per second (fps), 120fps, 240fps, etc., depending on the embodiment. In at least one embodiment, the camera may be able to use a rolling shutter, a global shutter, another type of shutter, or a combination thereof. In at least one embodiment, the color filter array may include a red transparent transparent ("RCCC") filter array, a red transparent transparent blue ("RCCB") filter array, a red blue green transparent ("RBGC") filter array, a Foveon X3 filter array, a Bayer sensor ("RGGB") filter array, a monochrome sensor filter array, and / or other types of filter arrays. In at least one embodiment, a clear pixel camera, such as one having an RCCC, RCCB, and / or RBGC color filter array, may be used in an effort to increase photosensitivity.

[0264] In at least one embodiment, one or more cameras can be used to perform advanced driver assistance system ("ADAS") functions (e.g., as part of a redundant or fail-safe design). For example, in at least one embodiment, a multi-function mono camera can be installed to provide functions including lane departure warning, traffic sign assistance, and intelligent headlight control. In at least one embodiment, one or more cameras (e.g., all cameras) can simultaneously record and provide image data (e.g., video).

[0265] In at least one embodiment, one or more cameras can be mounted in a mounting assembly, such as a custom designed (three-dimensional ("3D") printed) assembly, so as to cut out stray light and reflections from within the vehicle 1600 (e.g., reflections of the dashboard reflecting in the windshield mirror) that may interfere with the camera's image data capture capabilities. With respect to the rearview mirror mounting assembly, in at least one embodiment, the rearview mirror assembly can be 3D printed custom so that the camera mounting plate matches the shape of the rearview mirror. In at least one embodiment, one or more cameras can be integrated into the rearview mirror. In at least one embodiment, for side view cameras, one or more cameras can also be integrated within four pillars at each corner of the cabin.

[0266] In at least one embodiment, a camera (e.g., a forward-facing camera) having a field of view that includes a portion of the environment in front of the vehicle 1600 can be used for surround vision, as well as to help identify the forward path and obstacles with the assistance of one or more controllers 1636 and / or control SoCs, thereby providing information critical to generating an occupancy grid and / or determining a preferred vehicle path. In at least one embodiment, the forward-facing camera can be used to perform many ADAS functions similar to LIDAR, including but not limited to emergency braking, pedestrian detection, and collision avoidance. In at least one embodiment, the forward-facing camera can also be used for ADAS functions and systems, including but not limited to lane departure warning ("LDW"), automatic cruise control ("ACC"), and / or other functions (e.g., traffic sign recognition).

[0267] In at least one embodiment, a variety of cameras can be used in a forward-facing configuration, including, for example, a monocular camera platform including a CMOS ("complementary metal oxide semiconductor") color imager. In at least one embodiment, a wide-angle camera 1670 can be used to sense objects entering from the periphery (e.g., pedestrians, people crossing the street, or bicycles). Although in Fig. 16BOnly one wide-angle camera 1670 is shown, however, in other embodiments, there may be any number (including zero) of wide-angle cameras on the vehicle 1600. In at least one embodiment, any number of remote cameras 1698 (e.g., a remote stereo camera pair) may be used for depth-based object detection, particularly for objects for which a neural network has not yet been trained. In at least one embodiment, the remote cameras 1698 may also be used for object detection and classification and basic object tracking.

[0268] In at least one embodiment, any number of stereo cameras 1668 may also be included in the forward configuration. In at least one embodiment, one or more stereo cameras 1668 may include an integrated control unit including a scalable processing unit that may provide programmable logic ("FPGA") and a multi-core microprocessor with a controller area network ("CAN") or Ethernet interface integrated on a single chip. In at least one embodiment, such a unit may be used to generate a 3D map of the environment of the vehicle 1600, including distance estimates for all points in the image. In at least one embodiment, one or more stereo cameras 1668 may include, but are not limited to, a compact stereo vision sensor, which may include, but are not limited to, two camera lenses (one on each side) and an image processing chip that may measure the distance from the vehicle 1600 to the target object and use the generated information (e.g., metadata) to activate autonomous emergency braking and lane departure warning functions. In at least one embodiment, other types of stereo cameras 1668 may be used in addition to those described herein.

[0269] In at least one embodiment, a camera having a field of view of a portion of the environment including the sides of the vehicle 1600 (e.g., a side view camera) can be used for surround viewing to provide information for creating and updating occupancy grids and generating side collision warnings. For example, in at least one embodiment, surround cameras 1674 (e.g., Fig. 16B The vehicle 1600 may be provided with a plurality of surround cameras 1674 (e.g., four surround cameras shown) that can be positioned on the vehicle 1600. In at least one embodiment, the one or more surround cameras 1674 can include, but are not limited to, any number and combination of wide angle cameras, one or more fisheye cameras, one or more 360 ​​degree cameras, and / or the like. For example, in at least one embodiment, the four fisheye cameras can be located at the front, rear, and sides of the vehicle 1600. In at least one embodiment, the vehicle 1600 can use three surround cameras 1674 (e.g., left, right, and rear), and can utilize one or more other cameras (e.g., a forward-facing camera) as a fourth surround camera.

[0270] In at least one embodiment, a camera having a field of view that includes a portion of the environment behind the vehicle 1600 (e.g., a rear view camera) can be used for parking assistance, surround view, rear collision warning, and creating and updating occupancy grids. In at least one embodiment, a variety of cameras can be used, including but not limited to cameras that are also suitable as one or more forward-facing cameras (e.g., long-range camera 1698 and / or one or more mid-range cameras 1676, one or more stereo cameras 1668, one or more infrared cameras 1672, etc.), as described herein.

[0271] Reasoning and / or training logic 715 is used to perform reasoning and / or training operations associated with one or more embodiments. Fig. 7A 7B provides details about the reasoning and / or training logic 715. In at least one embodiment, the reasoning and / or training logic 715 may be Fig. 16B for use in a system for inferring or predicting operations based at least in part on weight parameters computed using neural network training operations, neural network functions and / or architectures, or neural network use cases described herein.

[0272] Such a component may be used to generate a single consistent tokenized description of at least a portion of a physical environment based in part on a set of observations and aligned map data.

[0273] Fig. 16C According to at least one embodiment, Fig.16A A block diagram of an example system architecture of an autonomous vehicle 1600. In at least one embodiment, Fig. 16C Each of one or more components, one or more features, and one or more systems of the vehicle 1600 in FIG. 1 is shown as being connected via a bus 1602. In at least one embodiment, the bus 1602 may include, but is not limited to, a CAN data interface (referred to herein alternatively as a "CAN bus"). In at least one embodiment, the CAN may be a network inside the vehicle 1600 that helps control various features and functions of the vehicle 1600, such as actuation of brakes, acceleration, braking, steering, wipers, etc. In one embodiment, the bus 1602 may be configured to have dozens or even hundreds of nodes, each with its own unique identifier (e.g., a CAN ID). In at least one embodiment, the bus 1602 may be read to find steering wheel angle, ground speed, engine revolutions per minute ("RPM"), button position, and / or other vehicle status indicators. In at least one embodiment, the bus 1602 may be a CAN bus that complies with ASIL B.

[0274] In at least one embodiment, in addition to or from CAN, FlexRay and / or Ethernet protocols may be used. In at least one embodiment, there may be any number of buses forming bus 1602, which may include, but are not limited to, zero or more CAN buses, zero or more FlexRay buses, zero or more Ethernet buses, and / or zero or more other types of buses using other protocols. In at least one embodiment, two or more buses may be used to perform different functions, and / or may be used for redundancy. For example, a first bus may be used for a collision avoidance function, and a second bus may be used for actuation control. In at least one embodiment, each of the buses 1602 may communicate with any component of the vehicle 1600, and two or more of the buses 1602 may communicate with corresponding components. In at least one embodiment, each of any number of systems on chip ("SoC") 1604 (such as SoC (A) and SoC (B)), each of one or more controllers 1636, and / or each computer within the vehicle may access the same input data (e.g., input from sensors of the vehicle 1600), and may be connected to a common bus, such as a CAN bus.

[0275] In at least one embodiment, the vehicle 1600 may include one or more controllers 1636, such as those described herein with respect to Fig.16A Those described. In at least one embodiment, the controller 1636 can be used for a variety of functions. In at least one embodiment, the controller 1636 can be coupled to any of the various other components and systems of the vehicle 1600 and can be used to control the vehicle 1600, the artificial intelligence of the vehicle 1600, the infotainment and / or other functions of the vehicle 1600.

[0276] In at least one embodiment, the vehicle 1600 may include any number of SoCs 1604. In at least one embodiment, each of the SoCs 1604 may include, but is not limited to, a central processing unit ("CPU(s)") 1606, a graphics processing unit ("GPU(s")) 1608, one or more processors 1610, one or more caches 1612, one or more accelerators 1614, one or more data stores 1616, and / or other components and features not shown. In at least one embodiment, the one or more SoCs 1604 may be used to control the vehicle 1600 in a variety of platforms and systems. For example, in at least one embodiment, the one or more SoCs 1604 may be combined with a high-definition ("HD") map 1622 in a system (e.g., a system of the vehicle 1600), which may be downloaded from one or more servers (e.g., a system of the vehicle 1600) via a network interface 1624. Fig. 16C) to obtain map refreshes and / or updates.

[0277] In at least one embodiment, one or more CPUs 1606 may include a CPU cluster or CPU complex (referred to herein alternatively as a "CCPLEX"). In at least one embodiment, one or more CPUs 1606 may include multiple cores and / or a secondary ("L2") cache. For example, in at least one embodiment, one or more CPUs 1606 may include eight cores in a mutually coupled multiprocessor configuration. In at least one embodiment, one or more CPUs 1606 may include four dual-core clusters, each with a dedicated L2 cache (e.g., a 2 megabyte (MB) L2 cache). In at least one embodiment, one or more CPUs 1606 (e.g., CCPLEX) may be configured to support simultaneous cluster operations such that any combination of clusters of one or more CPUs 1606 may be active at any given time.

[0278] In at least one embodiment, one or more CPUs 1606 may implement power management functions including, but not limited to, one or more of the following features: each hardware module may be automatically clock gated when idle to save dynamic power; each core clock may be gated when such core is not actively executing instructions due to executing a wait for interrupt ("WFI") / wait for event ("WFE") instruction; each core may be independently powered; each core cluster may be independently clock gated when all cores are clock gated or power gated; and / or each core cluster may be independently power gated when all cores are power gated. In at least one embodiment, one or more CPUs 1606 may further implement an enhanced algorithm for managing power states, wherein allowed power states and expected wakeup times are specified, and hardware / microcode determines the optimal power state...

Claims

1. A method comprising: Obtain a set of observations corresponding to the physical environment; identifying local map data aligned with the set of observations; as well as Generate a tokenized description of at least a portion of the environment using the trained language model and based at least on at least a subset of the set of observations and the local map data, and perform one or more operations corresponding to a machine based at least on the tokenized description.

2. The method of claim 1, further comprising: capturing sensor data corresponding to a physical environment of the machine using one or more sensors of the machine; extracting domain-related features from the sensor data using a perception module; as well as The set of observations corresponding to the physical environment is generated.

3. The method of claim 2, wherein the set of observations and the local map data are determined in relation to a current position of the machine in the physical environment, and wherein the one or more operations correspond to at least one of planning, control or navigation.

4. The method of claim 1, further comprising: determining an approximate reference location in the physical environment; selecting map data including the approximate reference location; as well as The map data is associated with the set of observations to identify the local map data that is aligned with the set of observations.

5. The method of claim 4, further comprising: The map data is associated with the set of observations using the trained model.

6. The method of claim 1 , wherein the trained language model fuses a portion of the set of observations with the local map data to generate a single consistent representation of the portion of the environment, and wherein the tokenized description is generated based in part on the single consistent representation.

7. The method of claim 1, wherein the one or more operations are domain-specific, and wherein the tokenized description is generated in a domain-specific language corresponding to the domain.

8. The method of claim 1, wherein the tokenized description includes tags for static objects and dynamic objects in the physical environment.

9. The method of claim 1, wherein the language model generates the tokenized description using incomplete local map data or an incomplete set of observations of the portion of the environment.

10. The method of claim 1, wherein the tokenized description is written in a road topology language RTL or other domain specific language DSL.

11. The method of claim 1, wherein the tokenized description is determined based on at least one of semantic, topological, geometric, kinematic, or relational information of features in the set of observations.

12. A processor, comprising: One or more circuits for: Generate a set of observations corresponding to the physical environment; identifying local map data aligned with the set of observations; as well as Processing data including at least a subset of the set of observations and the local map data based at least on a trained language model generates a tokenized description of at least a portion of the environment.

13. A processor as claimed in claim 12, wherein the trained language model is used to fuse a portion of the set of observations with the local map data to generate a single consistent representation of the portion of the environment, and wherein the tokenized description is generated based in part on the single consistent representation.

14. The processor of claim 12, wherein the tokenized description is to be used to perform domain-specific operations, and wherein the tokenized description is generated in a domain-specific language corresponding to the domain.

15. The processor of claim 12, wherein the trained language model is used to generate the tokenized description using incomplete local map data or an incomplete set of observations of the portion of the environment.

16. The processor of claim 12, wherein the processor is included in at least one of: A system for performing simulation operations; Systems for performing simulated operations to test or validate autonomous machine applications; Systems for performing digital twin operations; A system for performing light transport simulations; Systems for rendering graphical output; Systems for performing deep learning operations; A system for performing generative AI operations using a large language model (LLM); Systems implemented using edge devices; Systems for generating or presenting virtual reality (VR) content; Systems for generating or presenting augmented reality (AR) content; Systems for generating or presenting mixed reality (MR) content; A system comprising one or more virtual machines VM; A system implemented at least in part in a data center; A system for performing hardware testing using simulation; A system for performing generative operations using a language model LM; Systems for synthetic data generation; A collaborative content creation platform for 3D assets; or A system implemented at least in part using cloud computing resources.

17. A system comprising: One or more processors for generating a tokenized description of at least a portion of a physical environment using a trained language model based at least on a set of observations corresponding to the physical environment and local map data aligned with the set of observations.

18. A system as claimed in claim 17, wherein the trained language model fuses at least a portion of the set of observations with the local map data to generate a single consistent representation of the portion of the environment, and wherein the tokenized description is generated at least in part based on the single consistent representation.

19. The system of claim 17, wherein the tokenized description comprises a sequence of tokens indicating spatial information and semantic information of one or more static or dynamic objects identified in the physical location.

20. The system of claim 17, wherein the system comprises at least one of: A system for performing simulation operations; Systems for performing simulated operations to test or validate autonomous machine applications; Systems for performing digital twin operations; A system for performing light transport simulations; Systems for rendering graphical output; Systems for performing deep learning operations; A system for performing generative AI operations using a large language model (LLM); a system implemented using an edge device; Systems for generating or presenting virtual reality (VR) content; systems for generating or presenting augmented reality (AR) content; systems for generating or presenting mixed reality (MR) content; A system comprising one or more virtual machines VM; A system implemented at least in part in a data center; A system for performing hardware testing using simulation; A system for performing generative operations using a language model LM; A system for synthetic data generation; A collaborative content creation platform for 3D assets; or A system implemented at least in part using cloud computing resources.

Images