SYSTEM AND METHOD OF CAUSAL COMPOSITION DIFFUSION FOR TRAFFIC GENERATION IN A CLOSED REGULATOR

Abstract

A procedure comprises: receiving initial conditions for a traffic scenario encompassing a multitude of interacting agents, where the initial conditions define the states of the multitude of agents; identifying a causal structure among the multitude of agents based on their states; and ranking the multitude of agents based on the identified causal structure to determine a subset of key agents that have the greatest influence with respect to a controllability goal. For each agent in the multitude, the procedure comprises: generating a future trajectory using a reverse-sampling procedure of a diffusion model; and performing the reverse-sampling procedure by selectively applying a gradient of the controllability goal only to the identified subset of key agents.

Description

CROSS-REFERENCE TO RELATED REGISTRATION

[0001] This application claims priority under 35 USC §119(e) from the US Provisional Application filed on November 13, 2024, with serial number 63 / 720,114. The disclosure of that earlier application is incorporated into the disclosure of this application and is hereby included in its entirety by reference. INTRODUCTION

[0002] The information in this section serves to present the general context of the disclosure. Works of the inventors mentioned herein, insofar as they are described in this section, as well as aspects of the description that may not have been prior art at the time of filing, are neither expressly nor implicitly admitted as prior art against the present disclosure.

[0003] This disclosure relates generally to computer-implemented simulations and, in particular, to systems and methods for generating realistic and controllable traffic scenarios for testing and validating autonomous vehicles (AVs). The development and validation of safe and reliable AVs depends heavily on rigorous testing in a wide variety of driving scenarios. While real-world testing is essential, relying on it to cover the large number of potential interactions, especially rare but safety-critical long-tail events, is impractical and often dangerous. Consequently, realistic simulation has become an indispensable tool for evaluating AV performance.

[0004] An effective traffic simulator must generate scenarios that are both realistic and controllable. Realism ensures that the simulated behaviors of actors or agents (e.g., other cars, pedestrians) in the environment accurately reflect the complex, nuanced, and often unpredictable nature of interactions in the real world. Controllability allows developers and testers to deliberately create and analyze challenging situations, such as forcing a near miss or collision, to systematically explore the limits of an AV's capabilities. However, existing approaches to traffic simulation struggle to strike an appropriate balance between these two often competing objectives.Data-driven methods that learn from large datasets of real-world driving behavior can produce realistic general behaviors, but are often unable to generate new, safety-critical scenarios that are rare in the training data. Furthermore, when used in a closed-loop control system where the simulation evolves over time, these models can suffer from cumulative errors that cause the simulation to drift into unrealistic states.

[0005] Conversely, while rule-based approaches offer precise control, they often result in behaviors that feel scripted, rigid, and unrealistic because they fail to capture the adaptive decision-making of human drivers. Newer deep generative models, including diffusion models, have shown promise but still face a fundamental challenge: a conflict between the goals of realism and controllability. To steer a simulation toward a specific, user-defined outcome (e.g., a collision), it is often necessary to generate agent behaviors that are improbable and deviate significantly from realistic, data-learned patterns. This “gradient conflict” means that increasing controllability often comes directly at the expense of realism, and vice versa.Therefore, there is a need for a traffic scenario generation system that resolves this conflict and enables the creation of scenarios that are both highly realistic and precisely controllable, especially for generating safety-critical events. SUMMARY

[0006] One aspect of the disclosure provides a computer-implemented random composition diffusion method for closed-loop traffic generation, which, when executed on data processing hardware, causes the data processing hardware to perform operations that include: receiving initial conditions for a traffic scenario comprising a multitude of interacting agents, where the initial conditions define the states of the multitude of agents; and identifying a causal structure among the multitude of agents based on the states of the multitude of agents, where the causal structure defines causal influences between agents of the multitude. The operations also include: ranking the multitude of agents based on the identified causal structure to determine a subset of key agents that have the greatest influence with respect to a controllability goal.For each agent of the multitude of agents, the operations further include: generating a future trajectory using a reverse-sampling procedure of a diffusion model, and conducting the reverse-sampling procedure by selectively applying a gradient of the controllability target only to the identified subset of key agents, while the conduct for the remaining agents of the multitude of agents is determined based on the identified causal structure, thereby generating a final traffic scenario that satisfies the controllability target while remaining realistic.

[0007] Implementations of the revelation may include one or more of the following optional features. In some implementations, identifying the causal structure involves creating a decision causal graph (DCG), where the nodes of the DCG represent the agents and the lines represent the causal dependencies for future actions. In these implementations, the DCG may be generated using a scene encoder with a factored attention mechanism, identifying causal connections based on attention weights and / or kinematic factors between the plural of agents. Here, the kinematic factors may include a value for the time-to-collision (TTC) between pairs of agents from the plural.

[0008] In some examples, classifying the multitude of agents involves performing a graph-based analysis of the identified causal structure to determine the degree of interactivity for each agent in the multitude. In some implementations, conducting the reverse sampling procedure further involves applying a classifier-free guidance component. Here, the classifier-free guidance component comprises a weighted combination of an unconditional distribution based on an agent's own history and an intervened distribution based on the agent's causal parents, as defined by the causal structure.

[0009] In some examples, the controllability goal is linked to generating a safety-critical event. In these examples, the safety-critical event might include a collision between at least two agents, an off-road event for at least one agent, or a near-collision. In some implementations, the diffusion model is formulated as a constrained optimization problem within a constrained factorized Markov decision process (CFMDP). Here, the controllability goal is maximized subject to a reality requirement. In some examples, the diffusion model is a probabilistic denoising-diffusion model (DDPM), and the reverse-sampling procedure iteratively denoises a noise vector to generate each of the future trajectories.

[0010] Another aspect of the disclosure provides a system for random composition diffusion for closed-loop traffic generation, comprising data processing hardware and storage hardware communicating with the data processing hardware. The storage hardware stores instructions which, when executed on the data processing hardware, cause the data processing hardware to perform operations that include: receiving initial conditions for a traffic scenario with a multitude of interacting agents, where the initial conditions define the states of the multitude of agents; and identifying a causal structure among the multitude of agents based on the states of the multitude of agents, where the causal structure defines causal influences between agents of the multitude.The operations also include: ranking the multitude of agents based on the identified causal structure to determine a subset of key agents that have the greatest influence with respect to a controllability goal. For each agent in the multitude, the operations further include: generating a future trajectory using a reverse-sampling procedure of a diffusion model, and conducting the reverse-sampling procedure by selectively applying a gradient of the controllability goal only to the identified subset of key agents, while conducting for the remaining agents in the multitude is determined based on the identified causal structure, thereby producing a final traffic scenario that satisfies the controllability goal while remaining realistic.

[0011] This aspect can include one or more of the following optional features. In some implementations, identifying the causal structure involves creating a decision causal graph (DCG), where the nodes of the DCG represent the agents and the lines represent the causal dependencies for future actions. In these implementations, the DCG can be generated using a scene encoder with a factorized attention mechanism, identifying causal connections based on attention weights and / or kinematic factors between the set of agents. Here, kinematic factors can include a time-to-collision (TTC) value between pairs of agents from the set of agents.

[0012] In some examples, the classification of the multitude of agents includes: performing a graph-based analysis of the identified causal structure to determine the degree of interactivity for each agent in the multitude. In some implementations, conducting the reverse sampling procedure further includes: applying a classifier-free guiding component. Here, the classifier-free guiding component comprises a weighted combination of an unconditional distribution based on an agent's own history and an intervened distribution based on the agent's causal parents, as defined by the causal structure.

[0013] In some examples, the controllability goal is linked to the generation of a safety-critical event. In these examples, the safety-critical event might include a collision between at least two agents, an off-road condition for at least one agent, or a near-collision. In some implementations, the diffusion model is formulated as a constrained optimization problem within a constrained factorized Markov decision process (CFMDP). Here, the controllability goal is maximized subject to a reality requirement. In some examples, the diffusion model is a probabilistic denoising-diffusion model (DDPM), and the reverse-sampling procedure iteratively denoises a noise vector to generate each of the future trajectories.

[0014] One aspect of the disclosure provides a computer-implemented random composition diffusion procedure for closed-loop traffic generation. When executed on the data processing hardware, this procedure causes the hardware to perform operations that include: receiving initial conditions for a traffic scenario comprising a multitude of interacting agents, where the initial conditions define the states of the multitude of agents; and identifying a causal structure among the multitude of agents based on their states, where the causal structure defines causal influences between agents within the multitude. The operations also include: ranking the multitude of agents based on the identified causal structure to determine a subset of key agents that have the greatest influence with respect to a controllability goal.For each agent of the multitude of agents, the operations further include: generating a future trajectory using a reverse-sampling procedure of a diffusion model, whereby the reverse-sampling procedure is applied only to the identified subset of key agents by selectively applying a gradient of the controllability target.

[0015] The details of one or more implementations of the disclosure are set forth in the accompanying drawings and in the description below. Further aspects, features, and advantages will become apparent from the description and the drawings, as well as from the claims. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The drawings described here serve only to illustrate selected configurations and are not intended to limit the scope of this disclosure. Fig. Figure 1 is a schematic view of an exemplary system for random composition diffusion for traffic generation in a closed control loop. Fig. Figure 2 is a schematic view of exemplary components of a random composition diffusion model of the system of Fig. 1. Fig. Figure 3 is a schematic view of exemplary components of the random composition diffusion model of the system of Fig. 1. Fig. Figure 4 is a flowchart of an exemplary arrangement of operations for a causal composition diffusion procedure for traffic generation in a closed control loop. Fig. Figure 5 is a flowchart of an exemplary arrangement of operations for a causal composition diffusion procedure for traffic generation in a closed control loop.

[0017] The corresponding reference numbers denote the corresponding parts in the drawings. DETAILED DESCRIPTION

[0018] Exemplary configurations are now described in more detail with reference to the accompanying drawings. Exemplary configurations are provided so that this disclosure is thorough and conveys the full scope of the disclosure to those skilled in the art. Specific details are listed, such as examples of specific components, devices, and processes, to provide a thorough understanding of the configurations of this disclosure. It is clear to those skilled in the art that specific details need not be used, that exemplary configurations can be implemented in many different forms, and that the specific details and exemplary configurations should not be interpreted in such a way as to limit the scope of the disclosure.

[0019] The terminology used here serves only to describe certain exemplary configurations and is not intended to be restrictive. As used here, the singular articles "a," "an," and "the" can also include the plural forms unless the context clearly indicates otherwise. The terms "comprises," "comprehensive," "containing," and "exhibiting" are inclusive and therefore specify the presence of features, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, elements, components, and / or groups thereof. The procedural steps, processes, and operations described here are not to be interpreted as necessarily being carried out in the order discussed or presented, unless they are explicitly identified as such.Additional or alternative steps can be applied.

[0020] When an element or layer is described as "on," "engaging," "connected," "attached to," or "coupled" with another element or layer, it may be directly on, engaged, connected, attached, or coupled to that other element or layer, or there may be intervening elements or layers. Conversely, when an element is described as "directly on," "directly engaged with," "directly connected to," "directly attached to," or "directly coupled to" another element or layer, there must be no intervening elements or layers. Other words used to describe the relationship between elements should be interpreted similarly (e.g., "between" versus "directly between," "next to" versus "directly beside," etc.).As used here, the term “and / or” includes all combinations of one or more of the related listed elements.

[0021] The terms "first," "second," "third," etc., may be used here to describe different elements, components, areas, layers, and / or sections. These elements, components, areas, layers, and / or sections should not be restricted by these terms. These terms may only be used to distinguish one element, component, area, layer, or section from another. Terms such as "first," "second," and other numerical terms do not imply any sequence or order unless clearly indicated by the context. Thus, a first element, component, area, layer, or section discussed below could be referred to as a second element, component, area, layer, or section without deviating from the lessons of the exemplary configurations.

[0022] In this application, including the definitions below, the term "module" may be replaced by the term "circuit". The term "module" may refer to, be part of, or include: an application-specific integrated circuit (ASIC); a digital, analog, or mixed analog / digital discrete circuit; a digital, analog, or mixed analog / digital integrated circuit; a combinational logic circuit; a field-programmable gate array (FPGA); a processor (shared, dedicated, or group) that executes code; memory (shared, dedicated, or group) that stores the code executed by a processor; other suitable hardware components that provide the described functionality; or a combination of some or all of the above, e.g., in a system-on-a-chip.

[0023] The term "code," as used above, can include software, firmware, and / or microcode, and can refer to programs, routines, functions, classes, and / or objects. The term "shared processor" refers to a single processor that executes some or all of the code from multiple modules. The term "group processor" refers to a processor that, in combination with other processors, executes some or all of the code from one or more modules. The term "shared memory" refers to a single memory that stores some or all of the code from multiple modules. The term "group memory" refers to a memory that, in combination with other memories, stores some or all of the code from one or more modules. The term "memory" can be a subset of the term "computer-readable medium."The term "computer-readable medium" excludes transitory electrical and electromagnetic signals propagating through a medium and can therefore be considered tangible and non-transient storage. Non-restrictive examples of non-transient storage include tangible, computer-readable media, including non-volatile memory, magnetic storage, and optical storage.

[0024] The devices and methods described in this application can be implemented in whole or in part by one or more computer programs executed by one or more processors. The computer programs contain processor-executable instructions stored on at least one non-transitory, tangible, computer-readable medium. The computer programs may also contain and / or access stored data.

[0025] A software application (i.e., a software resource) can refer to computer software that causes a computing device to perform a task. In some examples, a software application may be called an "application," "app," or "program." Examples of applications include system diagnostics applications, system administration applications, system maintenance applications, word processing applications, spreadsheet applications, messaging applications, media streaming applications, social networking applications, and gaming applications.

[0026] Non-transitory memory can be physical devices used to temporarily or permanently store programs (e.g., instruction sequences) or data (e.g., program status information) for use by a computer. Non-transitory memory can be volatile and / or non-volatile addressable semiconductor memory. Examples of non-volatile memory include flash memory and read-only memory (ROM) / programmable read-only memory (PROM) / erasable programmable read-only memory (EPROM) / electronically erasable programmable read-only memory (EEPROM) (e.g., typically used for firmware such as boot programs). Examples of volatile memory include random access memory (RAM), dynamic random access memory (DRAM), static random access memory (SRAM), phase change memory (PCM), and floppy disks or tapes.

[0027] These computer programs (also referred to as programs, software, software applications or code) contain machine instructions for a programmable processor and may be implemented in a procedural and / or object-oriented high-level language and / or in assembly / machine language.

[0028] The terms "machine-readable medium" and "computer-readable medium" used herein refer to any computer program product, non-transient computer-readable medium, apparatus, and / or device (e.g., magnetic disks, optical disks, memory, programmable logic devices (PLDs)) that serves to provide machine instructions and / or data to a programmable processor, including a machine-readable medium that receives machine instructions as a machine-readable signal. The term "machine-readable signal" refers to any signal that serves to provide machine instructions and / or data to a programmable processor.

[0029] Various implementations of the systems and techniques described herein can be realized in digital electronic and / or optical circuits, integrated circuits, specially designed ASICs (application-specific integrated circuits), computer hardware, firmware, software, and / or combinations thereof. These various implementations can include implementation in one or more computer programs that are executable and / or interpretable on a programmable system comprising at least one programmable processor, which can be used for special or general purposes and is coupled such that it receives data and instructions from and transmits data and instructions to a storage system, as well as at least one input device and at least one output device.

[0030] The processes and logical sequences described in this description can be executed by one or more programmable processors, also known as data processing hardware, which run one or more computer programs to perform functions by responding to input data and producing outputs. The processes and logical sequences can also be executed by specialized logic circuits, such as an FPGA (Field Programmable Gate Array) or an ASIC (application-specific integrated circuit). Processors suitable for executing a computer program include, for example, both general-purpose and specialized microprocessors, as well as one or more processors from any type of digital computer. Generally, a processor receives instructions and data from read-only memory, random-access memory, or both.The essential elements of a computer are a processor for executing instructions and one or more storage devices for storing instructions and data. Generally, a computer also includes one or more mass storage devices for storing data, such as magnetic, magneto-optical, or optical disks, or is operationally connected to them to receive data from or transmit data to them. However, a computer does not necessarily have to have such devices. Computer-readable media suitable for storing computer program instructions and data include all forms of non-volatile memory, media, and storage devices, including, for example, semiconductor memory devices such as EPROM, EEPROM, and flash memory devices; magnetic disks, such as internal hard disks or removable media; magneto-optical disks; and CD-ROM and DVD-ROM disks.The processor and memory can be supplemented by special logic circuits or integrated into them.

[0031] To enable interaction with a user, one or more aspects of the revelation can be implemented on a computer that has a display device, such as a CRT (cathode ray tube), LCD (liquid crystal display), or touchscreen monitor to show information to the user, and optionally a keyboard and pointing device, such as a mouse or trackball, with which the user can input information into the computer. Other types of devices can also be used for user interaction; feedback to the user can be any form of sensory feedback, such as visual, auditory, or tactile feedback; and user input can be received in any form, including auditory, verbal, or tactile input.Furthermore, a computer can interact with a user by sending and receiving documents to a device used by the user; for example, by sending web pages to a web browser on a user's client device in response to requests received from the web browser.

[0032] The development and validation of safe and reliable AVs depend heavily on rigorous testing in a wide variety of driving scenarios. While real-world testing is essential, relying on it to cover the vast number of potential interactions, especially rare but safety-critical long-tail events, is impractical and often dangerous. Consequently, realistic simulation has become an indispensable tool for evaluating AV performance. An effective traffic simulator must generate scenarios that are both realistic and controllable. Realism ensures that the simulated behaviors of actors or agents (e.g., other cars, pedestrians) in the environment accurately reflect the complex, nuanced, and often unpredictable nature of real-world interactions.Controllability allows developers and testers to deliberately create and analyze challenging situations, such as forcing a near miss or collision, to systematically explore the limits of an AV's capabilities. However, existing traffic simulation approaches struggle to strike an appropriate balance between these two often competing objectives. Data-driven methods that learn from large datasets of real-world driving behavior can produce realistic general behaviors, but are often unable to generate novel, safety-critical scenarios that are rare in the training data. Furthermore, when used in a closed-loop system where the simulation evolves over time, these models can suffer from cumulative errors that cause the simulation to drift into unrealistic states.

[0033] Conversely, while rule-based approaches offer precise control, they often result in behaviors that feel scripted, rigid, and unrealistic because they fail to capture the adaptive decision-making of human drivers. Newer deep generative models, including diffusion models, have shown promise but still face a fundamental challenge: a conflict between the goals of realism and controllability. To steer a simulation toward a specific, user-defined outcome (e.g., a collision), it is often necessary to generate agent behaviors that are improbable and deviate significantly from realistic, data-learned patterns. This “gradient conflict” means that increasing controllability often comes directly at the expense of realism, and vice versa.Therefore, there is a need for a traffic scenario generation system that resolves this conflict and enables the creation of scenarios that are both highly realistic and precisely controllable, especially for generating safety-critical events.

[0034] The present system and method generate realistic and controllable traffic scenarios in a closed-loop control system, overcoming the aforementioned limitations of the state of the art. The system and method incorporate a diffusion model, referred to as the Causal Composition Diffusion model (CCDiff), which resolves the inherent conflict between realism and controllability by identifying and utilizing the underlying causal structure of interactions within a traffic scene. The method formulates the scenario generation task as a constrained optimization problem that aims to maximize a user-defined controllability goal (e.g., triggering a safety-critical event) while simultaneously fulfilling a realism requirement. At its core, the system uses a diffusion model to generate agent trajectories.The crucial innovation lies in the way this generation process is conducted.

[0035] First, a causal logic module analyzes the scene to automatically generate a decision causal graph (DCG) that maps the cause-and-effect relationships between agents. Based on this graph, the agents are ranked according to their influence and interactivity, identifying a small subset of key agents crucial for achieving the desired outcome. Second, a novel causal composition guidance mechanism is used to control the CCDiff generation process. This guidance is structured and selective. A gradient related to the goal of controllability is applied only to the identified key agents. This targeted intervention efficiently steers the scenario toward the desired outcome.Simultaneously, the behavior of all other agents is guided by the identified causal structure to ensure that their actions remain consistent and realistic within the context of the scene. This decoupling of guidance allows the system to avoid the gradient conflicts that plague previous methods, thus achieving a better balance between realism and controllability.

[0036] In Fig. Figure 1 represents a system 100 for the random distribution of the composition for traffic generation in a closed-loop control system. System 100 comprises a remote computer system 50 and an autonomous agent. Although the autonomous agent is represented as vehicle 10, the systems and procedures described here are broadly applicable to other types of autonomous agents. Such agents can include, among others, autonomous mobile robots (AMRs) working in warehouses, robotic manipulators performing tasks in dynamic environments, unmanned aerial vehicles (UAVs), or agricultural and construction machinery. Furthermore, the principles can be applied to simulation systems for modeling agent behavior, for example, in air traffic control systems or in pedestrian flow analysis.The remote computer system 50 can be a single computer, multiple computers, or a distributed system, such as a cloud computing environment, with data processing hardware 52 and memory 54. The memory 54 stores instructions which, when executed by the data processing hardware 52, configure the remote computer system 50 to operate as a causal composition diffusion model 200. The causal composition diffusion model 200 is configured to generate a scenario for a future trajectory 234, which is then used to validate a driving model 120 executed by the onboard driver assistance system 12. In this case, the future trajectory 234 can form a final traffic scenario 232 that fulfills a controllability goal of the causal composition diffusion model 200. As used here, the controllability goal can be the generation orThe generation of a safety-critical event includes, for example, a collision, an off-road event, a near miss, or an overspeed event. Even though it is a remote system, the functionality of the causal composition diffusion model 200 can, in some implementations, be executed wholly or partially on computer resources within the vehicle 10.

[0037] After the causal composition diffusion model 200 has generated the future trajectory 234, it makes this trajectory available for testing and validating an inference task performed by the driving model 120. The future trajectory 234 can be forwarded to a mobile platform, such as the vehicle 10 shown, for execution by an onboard controller 14. The presented methods also extend beyond perception and control tasks. The vehicle controller 14 is part of an onboard control system, such as the driver assistance system 12 shown, which also includes an onboard computing system 30 with its own data processing hardware 32 and memory 34, a sensor system 20, a user interface system 40, and a network interface (not shown).The vehicle controller 14 uses the driving model 120 to perform an inference task, which involves processing real-time sensor data from the sensor system 20. The sensor system 20 can contain various sensors, such as one or more cameras 22, radar sensors 24, or lidar sensors 26. The result of the inference task is provided to one or more functions of the driver assistance system 12, such as an adaptive cruise control system or an automatic emergency braking system.

[0038] As in the Fig. 2 and Fig. As shown in Figure 3, the causal composition diffusion model 200 comprises a scene encoder 210, a causal logic 300, a ranking module 220, and a diffusion model 230. The scene encoder 210 is configured to receive as input the history 208 of previous traffic scenarios and performs structured scene coding to encode the history 208 and generate as output a predicted action 212 of the history 208. The causal composition diffusion model 300 is configured to receive as input the initial conditions 202 for a traffic scenario with a multitude of interacting agents 204. In this case, the initial conditions 202 define states 206 of each of the multitude of interacting agents 204. The causal composition diffusion model 300 is configured to identify a random structure 332 among the multitude of agents 204 based on the states 206 of the multitude of agents 204.

[0039] As in Fig. As shown in Figure 2, the causal logic 300 produces as output the random structure 332, which defines the random influences between the agents 204 of the plurality of agents 204. In some implementations, the random structure 332 comprises a decision causal graph (DCG) with a plurality of nodes representing agents 204 and edges or lines representing causal dependencies for future actions. In some cases, the DCG of the random structure 332 is generated using the scene encoder 210 with a factorized attention mechanism. Here, the random connections can be determined based on attention weights or kinematic factors between the plurality of agents 204. These kinematic factors can include a value for the time-to-collision (TTC) between pairs of agents 204 from the plurality of agents 204.

[0040] With reference to Fig. 3. The causal logic 300 can include a tokenizer 310, an attention layer 320, and a masking module 330, which work together to generate the random structure 332 (e.g., a factored DCG). Here, the causal logic 300 encodes the movement sequences 208 of different agents 204 in the initial conditions 202 based on spatial attention and then discovers the DCG based on the factored attention masks and kinematic factors. Finally, the causal logic optimizes its controllability by masking the non-key agents 204 in order to control the reverse sampling process of the diffusion model 230 in a structured manner. The Tokenizer 310 comprises a transformer-based structure configured to receive the initial conditions 202 and the history 208 as input and to embed the history 208 of the agents 204 to generate an agent embedding 312.Here, the agent embedding 312 encompasses the history or course of each agent 204 in relation to the course 208 of all other agents 204 in the initial conditions 202. To facilitate relational inferences, both the absolute and relative characteristics of the agents 204 are included, including the position, velocity, distance, and time to collision (TTC) of each agent 204 relative to the other agents 204.

[0041] The attention layer 320 then aggregates all temporal information from the agent embedding 312 to generate an attention output 322. To discover further useful spatial parent-child relationships, the causal logic 300 applies a two-stage causal inference to identify the DCG 332 in the spatiotemporal interaction of the agents 204. First, by setting a hard condition on the perceptual field of the neighborhood, by severing the unnecessary causal link between the agents' states 206 and the corresponding actions at time step t. Second, by applying the first tunable hard condition as a memory mask to the agents' attention weights 204 via the masking module 330, as follows: Gij(τt)=Mij(τt)⋅softmax((qWqhtT)(kWkhtij)dk)

[0042] Here, M denotes the memory mask extracted with relative TTC features. For each agent 204 in the environment, a threshold value C of DCG 332 is assigned in the initial conditions 202 as follows: Mij(τt)={1,fTTC(Øothers(stj→i)≤CTTC)0,otherwise

[0043] The masking module 330 can tune the threshold C to control the rarity of the final DCGs 332 so that the diffusion model 230 aggregates the map information c and the state of random parent agents 204 to obtain a final action for the future trajectory 234 of agent 204.

[0044] The ranking module 220 is configured to receive as input the random structure 332 containing the multitude of agents 204 and their respective causal influences. It then ranks the multitude of agents 204 to determine a subset of key agents 204K that are most influential with respect to a controllability goal of the causal composition diffusion model 200. In other words, the ranking module 220 performs a top-K ranking of the most influential agents 204 in the random structure 332 to identify the subset of key agents 204K. In some cases, the ranking module 220 performs a graph-based analysis of the identified random structure 332 to determine the degree of interactivity for each agent 204 in the multitude of agents 204. The ranking module 220 can generate as output the subset of key agents 204K for the diffusion model 230.

[0045] The diffusion model 230 then receives as input the random structure 332, which contains the agents 204 and the subset of key agents 204K, and the predicted action 212 generated by the scene encoder 210, and produces as output a future trajectory 232 for each agent 204 in the random structure 332. The diffusion model 230 can generate each future trajectory 232 using a reverse sampling procedure. In some cases, the diffusion model is a denoising diffusion probabilistic model (DDPM), where the reverse sampling procedure iteratively denoises a noise vector to generate each of the future trajectories 232 of the agents 204.In particular, the causal composition diffusion model 200 is configured to perform the reverse sampling procedure of the diffusion model 230 by selectively applying a gradient of the controllability target only to the specific subset of key agents 204K. Here, the guidance for the remaining agents 204 is determined based on the random structure 332 identified by the causal logic. By branching the guidance of the reverse sampling procedure, the diffusion model 230 generates a final traffic scenario 234 of all future trajectories 232 of the agents 204, which fulfills the controllability target while remaining realistic.

[0046] The causal composition diffusion model 200 can factor the goal of controllability by formulating a closed-loop traffic simulation as a Markov Decision Process (MDP) problem and a diffusion model 230 ( Fig. 2) used for sequential modeling to learn a controllable simulation policy. To exploit the causal structure between a state 206, an action, and a reward space, the causal composition diffusion model 200 is defined by a constrained factored MDP and a decision causal graph (i.e., a causal structure 332). The constrained factored MDP (CFMDP, or constrained factored MDP) is an MDP in which the state space S and a reward function R are factored to exploit the structure of the problem. A CFMDP is defined by the tuple: M F= (S, A, P, R, C, s0). Here, S denotes the factored state space, which represents a motion trajectory space at a current time step t for each agent 204, and A denotes the factored action space, which consists of interventions in the subsequent derivational behaviors for each agent 204 in the scenario. P denotes the joint transition dynamics, defined via the pairs of state S and action A, and defines the deterministic vehicle dynamics for each agent 204 in the environment. R denotes the reward target for collisions, off-road driving, speeding, or other targets, with each subset of R specifying the state factors that influence the reward. C denotes the requirement function, which specifies the degree of realism of the generated trajectories of the learned simulation policy with respect to the dataset policy, where a lower requirement value indicates a higher degree of realism.The initial state is denoted by s0, which lies in a factored state space s.

[0047] As mentioned above, for each time step t the causal structure 332 is defined as G, where G ij = 0 if and only if the future action of a particular agent 204 is conditionally independent of the history 208 of agent 204. If G 2,3 If = 1, the random structure 332 contains a causal line for this particular agent 204. Considering the above, the causal composition diffusion model 200 defines a set policy in which the causal parents of each agent 204 in the random structure 332 are used in the decision-making process to identify the random structure 332. Given the CFMDP and known vehicle dynamics, the causal composition diffusion model 200 factors the goal of optimal scenario generation in a closed-loop control system as follows: maxP(ϑt=1,τt|τt−1)↔maxP(ϑt=1|τt)P(τt|τt−1)↔maxPπ(ϑt=1|stat)π(a tst)P(st|st−1at−1)↔maxπ∏j=1drexp(Rj(stIjπ(st)))∏i=1Nπi(ati|st) where the first term corresponds to controllability (i.e., the probability of optimality specified by a user-defined reward target), and the second term corresponds to realism (i.e., the probability of the generated behavior in the future trajectory 232). The evaluation function of the maximum likelihood target can then be expressed as follows: ∇logP=∑j=1dr∇τRj(stIjπ(st))+∑i=1N∇τlogπi(ati|st)

[0048] Unlike normal scenarios where the optimization of imitation is essentially linked to the reward for rule compliance, in safety-critical leadership, R (3)This leads to gradient conflict. To resolve these gradient conflict problems, the causal composition diffusion model 200 prioritizes the agent index 204 and maximizes the reward while maintaining a high probability of the learned policies or strategies, i.e., a smaller reality gap between the learned policies and the behavioral policies. Here, the diffusion model 230 can use a Lagrange multiplier and a structured projected gradient descent to solve the constrained optimization with the following maximum likelihood estimation problem: maxπ∈,ρ∈{0,1}N,G∈{0,1}N×N∏j=1drexp(Rj(τtIj;ρ))∏i∈[N]ρi=1πi(ati|PAtG(i)) under the condition |G| ≤ C sparsity , Σ i ρ i ≤ N c , where the degree of reality is changed by altering the degree of restriction of N c , C sparsity ∈ Z + can be controlled.

[0049] Fig. Figure 4 shows a flowchart of an exemplary sequence of operations for a causal composition diffusion procedure 400 for traffic generation in a closed-loop control system. The procedure 400 can be described with reference to the Fig. 1-3 are described. The data processing hardware (e.g., data processing hardware 52 of Fig. 1) can execute instructions stored on the memory hardware (e.g., memory hardware 54 of Fig. 1) are stored to perform the exemplary arrangement of operations for procedure 400. In operation 402, procedure 400 includes receiving initial conditions 202 for a traffic scenario with a plurality of interacting agents 204, where the initial conditions 202 define the states 206 of the plurality of agents 204. In operation 404, procedure 400 includes identifying a causal structure 332 between the plurality of agents 204 based on the states 206 of the plurality of agents 204. The causal structure 332 defines causal influences between the agents 204 of the plurality of agents 204.

[0050] In Operation 406, Procedure 400 also includes classifying the multitude of agents 204 based on the identified causal structure 332 to determine a subset of key agents 204K that have the greatest influence with respect to a controllability goal. For each agent 204 from the multitude of agents 204, Procedure 400 also includes, in Operation 408, generating a future trajectory 232 using a reverse sampling process of a diffusion model 230. In Operation 410, Procedure 400 further includes performing the reverse sampling process by selectively applying a gradient of the controllability goal only to the determined subset of key agents 204K.Here, the guidance for the remaining agents 204 of the multitude of agents 204 is determined on the basis of the identified causal structure 332, thereby generating a final traffic scenario 234 that fulfills the goal of controllability while remaining realistic.

[0051] Fig. Figure 5 shows a flowchart of an exemplary sequence of operations for a causal composition diffusion procedure 500 for traffic generation in a closed-loop control system. The procedure 500 can be described with reference to the Fig. 1-3 are described. The data processing hardware (e.g., data processing hardware 52 of Fig. 1) can execute instructions stored on the memory hardware (e.g., memory hardware 54 of Fig.1) are stored to perform the exemplary arrangement of operations for procedure 500. In operation 502, procedure 500 includes receiving initial conditions 202 for a traffic scenario with a plurality of interacting agents 204, where the initial conditions 202 define the states 206 of the plurality of agents 204. In operation 504, procedure 500 includes identifying a causal structure 332 between the plurality of agents 204 based on the states 206 of the plurality of agents 204. The causal structure 332 defines causal influences between agents 204 of the plurality of agents 204.

[0052] In Operation 506, Procedure 500 also includes classifying the plurality of agents 504 based on the identified causal structure 332 to determine a subset of key agents 204K that have the greatest influence with respect to a controllability goal. For each agent 204 from the plurality of agents 204, Procedure 500 also includes, in Operation 508, generating a future trajectory 232 using a reverse sampling process of a diffusion model 230. In Operation 510, Procedure 500 further includes performing the reverse sampling process by selectively applying a gradient of the controllability goal only to the determined subset of key agents 204K.

[0053] Several implementations have been described. However, it goes without saying that various modifications can be made without deviating from the spirit and scope of the disclosure. Accordingly, other embodiments also fall within the scope of protection of the following claims.

[0054] The foregoing description serves for illustration and description purposes. It is not intended to be exhaustive or to limit the disclosure. Individual elements or features of a particular configuration are generally not restricted to that particular configuration but are optionally interchangeable and may be used in a selected configuration even if they are not specifically shown or described. The same may also be varied in many ways. Such variations are not to be considered outside the scope of disclosure, and all such modifications are to be included within the scope of protection of the disclosure. QUOTES INCLUDED IN THE DESCRIPTION

[0000] This list of documents cited by the applicant was automatically generated and is included solely for the reader's convenience. The list is not part of the German patent or utility model application. The DPMA accepts no liability for any errors or omissions. Cited patent literature

[0000] US 63 / 720,114

[0001]

Claims

[1] A computer-implemented method which, when executed on data processing hardware, causes the data processing hardware to perform operations, comprising: Receiving initial conditions for a traffic scenario, encompassing a multitude of interacting agents, where the initial conditions define the states of the multitude of agents; Identifying a causal structure between the multitude of agents based on the states of the multitude of agents, where the causal structure defines causal influences between agents of the multitude of agents; Classifying the multitude of agents based on the identified causal structure to determine a subset of key agents that have the greatest influence with respect to a controllability goal; For each agent in a multitude of agents: generating a future trajectory using a reverse-sampling method of a diffusion model; and The reverse sampling procedure is carried out by selectively applying a gradient of the controllability target only to the identified subset of key agents, while the leading for the remaining agents of the multitude of agents is determined based on the identified causal structure, thereby generating a final traffic scenario that satisfies the controllability target while remaining realistic. [2] Method according to claim 1, wherein identifying the causal structure comprises: creating a decision causal graph (DCG), wherein the nodes of the DCG represent the agents and the lines represent the causal dependencies for future actions. [3] Method according to claim 2, wherein the DCG is generated using a scene encoder with a factorized attention mechanism, and wherein causal links are identified between the plurality of agents based on attention weights and / or kinematic factors. [4] Method according to claim 3, wherein the kinematic factors comprise a value for the time to collision (TTC) between pairs of agents from the plurality of agents. [5] Method according to claim 1, wherein the classification of the plurality of agents comprises: performing a graph-based analysis of the identified causal structure to determine the degree of interactivity for each agent of the plurality of agents. [6] The method of claim 1, wherein the reverse sampling procedure further comprises: applying a classifier-free guiding component, wherein the classifier-free guiding component comprises: a weighted combination of an unconditional distribution based on an agent's own history and an intervened distribution based on the agent's causal parents as defined by the causal structure. [7] Method according to claim 1, wherein the controllability objective is linked to the generation of a safety-critical event. [8] Method according to claim 7, wherein the safety-critical event comprises: a collision between at least two agents, a departure from the road for at least one agent or a near-collision. [9] Method according to claim 1, wherein the method is formulated as a constrained optimization problem within the framework of a constrained factorized Markov decision process (CFMDP), wherein the controllability objective is maximized under the condition of realism. [10] Method according to claim 1, wherein the diffusion model is a probabilistic denoising diffusion model (DDPM) and the reverse sampling method iteratively denoises a noise vector to generate each of the future trajectories.

Images