System and method for causal combination diffusion for closed-loop traffic generation

Abstract

A method includes receiving an initial condition of a traffic scene including a plurality of interactive agents, the initial condition defining a state of the plurality of agents; identifying a causal structure between the plurality of agents based on the states of the plurality of agents; and ranking the plurality of agents based on the identified causal structure to determine a subset of key agents that have the greatest impact on the controllability goal. For each agent of the plurality of agents, the method includes generating a future trajectory using a reverse sampling process of a diffusion model, and directing the reverse sampling process by selectively applying a gradient of a controllability target only to a determined subset of key agents.

Description

[0001] Cross-reference to related applications

[0002] This application claims priority to U.S. Provisional Application Serial No. 63 / 720,114, filed November 13, 2024, pursuant to §119(e). The disclosure of the earlier application is considered part of the disclosure of this application and is incorporated herein by reference in its entirety.

[0003] introduce

[0004] The information provided in this section is for the purpose of presenting the general context of this disclosure. The work of the currently named inventors, to the extent described in this section, and in respect of aspects that may not otherwise qualify as prior art at the time of filing, is neither expressly nor implicitly acknowledged as prior art to this disclosure. Technical Field

[0005] This disclosure generally relates to computer-implemented simulations, and more specifically to systems and methods for generating realistic and controllable traffic scenarios for the testing and validation of autonomous vehicles (AVs). Background Technology

[0006] The development and validation of safe and reliable AVs rely heavily on rigorous testing across a wide range of driving scenarios. While real-world testing is essential, it is impractical and often dangerous to rely on it to cover the vast number of potential interactions, especially the rare but safety-critical "long-tail" events. Therefore, high-fidelity simulation has become an important tool for evaluating AV performance.

[0007] Effective traffic simulators must generate scenarios that are both realistic and controllable. Realism ensures that the simulated behavior of surrounding agents (e.g., other cars, pedestrians) accurately reflects the complex, subtle, and often unpredictable nature of real-world interactions. Controllability allows developers and testers to specifically create and analyze challenging situations, such as induced hazard or collision events, to systematically probe the limits of an AV's capabilities. However, existing traffic simulation methods struggle to adequately balance these two often competing goals. Data-driven approaches that learn from large datasets of real-world driving can produce realistic, common behaviors but often fail to generate novel, safety-critical scenarios rarely found in the training data. Furthermore, when used in closed-loop environments where simulations evolve over time, these models can suffer from compounded errors, causing the simulation to drift into unrealistic states.

[0008] Conversely, rule-based methods offer precise control but often produce seemingly rote, rigid, and unrealistic behavior because they fail to capture the adaptive decision-making of human drivers. Recent deep generative models, including diffusion models, have shown promise but still face a fundamental challenge: the conflict between realism and controllability. The process of guiding a simulation toward a specific, user-defined outcome (e.g., a collision) typically requires generating proxy behaviors that are unlikely and significantly deviate from realistic patterns learned from data. This “gradient conflict” means that increasing controllability often comes at the direct cost of sacrificing realism, and vice versa. Therefore, a traffic scene generation system is needed that can resolve this conflict, enabling the creation of scenarios that are both highly realistic and precisely controllable, particularly for safety-critical events. Summary of the Invention

[0009] One aspect of this disclosure provides a computer-implemented method for causal combinatorial diffusion in closed-loop traffic generation. When executed on data processing hardware, the method causes the hardware to perform operations including receiving initial conditions of a traffic scene comprising a plurality of interacting agents, the initial conditions defining the states of the agents, and identifying a causal structure among the agents based on the states of the agents, the causal structure defining causal influence among the agents. The operations further include ranking the agents based on the identified causal structure to determine a subset of key agents most influential with respect to a controllability objective. For each agent, the operations further include generating a future trajectory using a backsampling process of a diffusion model, and guiding the backsampling process by selectively applying the gradient of the controllability objective only to the determined subset of key agents, while determining guidance for the remaining agents based on the identified causal structure, thereby generating a final traffic scene that satisfies the controllability objective while maintaining realism.

[0010] Implementations of this disclosure may include one or more of the following optional features. In some implementations, identifying causal structures includes generating a decision causal graph (DCG), where nodes represent agents and edges represent causal dependencies of future actions. In these implementations, a scene encoder with a factorization attention mechanism can be used to generate the DCG, where causal connections are identified based on at least one of attention weights or kinematic factors among multiple agents. Here, kinematic factors may include collision time-to-collision (TTC) values ​​between pairs of agents among the multiple agents.

[0011] In some examples, ranking multiple agents includes performing graph-based analysis on the identified causal structures to determine the degree of interactivity of each agent among the multiple agents. In some implementations, the guided backsampling process also includes applying a classifier-free guided component. Here, the classifier-free guided component comprises a weighted combination of an unconditional distribution based on the agent's own history and an intervention distribution based on the agent's causal parents as defined by the causal structure.

[0012] In some examples, the controllability objective is associated with generating safety-critical events. These safety-critical events can include one of the following: a collision between at least two agents, an off-road event involving at least one agent, or a near-miss event. In some implementations, the diffusion model is formulated as a constrained optimization problem in a constrained factor Markov decision process (CFMDP). Here, the controllability objective is maximized under realistic constraints. In some examples, the diffusion model is a denoised diffusion probability model (DDPM), and the backsampling process iteratively denoises the noise vector to generate each future trajectory.

[0013] Another aspect of this disclosure provides a system for causal combination diffusion in closed-loop traffic generation, including data processing hardware and memory hardware communicating with the data processing hardware. The memory hardware stores instructions that, when executed by the data processing hardware, cause the data processing hardware to perform operations including receiving initial conditions of a traffic scene comprising a plurality of interacting agents, the initial conditions defining the states of the plurality of agents, and identifying a causal structure among the plurality of agents based on the states of the agents, the causal structure defining causal influence among the agents. The operation further includes ranking the plurality of agents based on the identified causal structure to determine a subset of key agents most influential with respect to a controllability objective. For each of the plurality of agents, the operation further includes generating a future trajectory using a backsampling process of a diffusion model, and guiding the backsampling process by selectively applying the gradient of the controllability objective only to the determined subset of key agents, while determining guidance for the remaining agents among the plurality of agents based on the identified causal structure, thereby generating a final traffic scene that satisfies the controllability objective while maintaining realism.

[0014] This aspect may include one or more of the following optional features. In some implementations, identifying causal structures includes generating a decision causal graph (DCG), where nodes represent agents and edges represent causal dependencies of future actions. In these implementations, a scene encoder with a factorization attention mechanism can be used to generate the DCG, where causal connections are identified based on at least one of attention weights or kinematic factors among multiple agents. Here, kinematic factors may include collision time-of-competition (TTC) values ​​between pairs of agents among the multiple agents.

[0015] In some examples, ranking multiple agents involves performing graph-based analysis on the identified causal structures to determine the degree of interactivity of each of the multiple agents. In some implementations, the guided backsampling process also includes applying a classifier-free guided component. Here, the classifier-free guided component comprises a weighted combination of an unconditional distribution based on the agent's own history and an intervention distribution based on the agent's causal parents as defined by the causal structure.

[0016] In some examples, the controllability objective is associated with generating safety-critical events. These safety-critical events may include a collision between at least two agents, an off-road incident involving at least one agent, or a near-miss incident. In some implementations, the diffusion model is formulated as a constrained optimization problem in a constrained factor Markov decision process (CFMDP). Here, the controllability objective is maximized under realistic constraints. In some examples, the diffusion model is a denoised diffusion probability model (DDPM), and the backsampling process iteratively denoises the noise vector to generate each future trajectory.

[0017] One aspect of this disclosure provides a computer-implemented method for causal combinatorial diffusion in closed-loop traffic generation. When executed on data processing hardware, the method causes the hardware to perform operations including receiving initial conditions of a traffic scenario comprising a plurality of interacting agents, the initial conditions defining states of the agents; and identifying a causal structure among the agents based on the states of the agents, the causal structure defining causal influence among the agents. The operations further include ranking the agents based on the identified causal structure to determine a subset of key agents most influential with respect to a controllability objective. For each agent, the operations further include generating a future trajectory using a backsampling process of a diffusion model, the backsampling process being guided by selectively applying the gradient of the controllability objective only to the determined subset of key agents.

[0018] Details of one or more embodiments of this disclosure are set forth in the accompanying drawings and the following description. Other aspects, features, and advantages will be apparent from the specification, drawings, and claims. Attached Figure Description

[0019] The accompanying drawings described herein are for illustrative purposes only for the selected configurations and are not intended to limit the scope of this disclosure.

[0020] Figure 1 This is a schematic diagram of an exemplary system for the causal combination diffusion of closed-loop traffic generation.

[0021] Figure 2 yes Figure 1A schematic diagram of an example component of a causal combination diffusion model of a system.

[0022] Figure 3 yes Figure 1 A schematic diagram of an example component of a causal combination diffusion model of a system.

[0023] Figure 4 This is a flowchart illustrating an exemplary arrangement of operations for a method of causal combination diffusion for closed-loop traffic generation.

[0024] Figure 5 This is a flowchart illustrating an exemplary arrangement of operations for a method of causal combination diffusion for closed-loop traffic generation.

[0025] Throughout the accompanying drawings, corresponding reference numerals indicate the relevant parts. Detailed Implementation

[0026] The exemplary configuration will now be described more fully with reference to the accompanying drawings. Exemplary configurations are provided so that this disclosure will be thorough and will fully communicate the scope of this disclosure to those skilled in the art. Specific details, such as examples of specific components, apparatus, and methods, are set forth to provide a thorough understanding of the configurations of this disclosure. It will be apparent to those skilled in the art that specific details are not required, that the exemplary configurations may be implemented in many different forms, and that the specific details and exemplary configurations should not be construed as limiting the scope of this disclosure.

[0027] The terminology used herein is for the purpose of describing particular exemplary configurations only and is not intended to be restrictive. As used herein, the singular articles “a,” “an,” and “the” may be intended to include plural forms as well, unless the context clearly indicates otherwise. The terms “comprises,” “comprising,” “including,” and “having” are inclusive and therefore specify the presence of features, steps, operations, elements, and / or components, but do not preclude the presence or addition of one or more other features, steps, operations, elements, components, and / or groups thereof. Unless specifically identified as an order of execution, the method steps, processes, and operations described herein should not be construed as requiring them to be performed in the specific order discussed or shown. Additional or alternative steps may be employed.

[0028] When an element or layer is referred to as being “on,” “joined to,” “connected to,” “attached to,” or “linked to” another element or layer, it may be directly on, joined to, connected to, attached to, or linked to the other element or layer, or there may be intermediate elements or layers present. Conversely, when an element is referred to as being “directly on,” “directly joined to,” “directly connected to,” “directly attached to,” or “directly linked to” another element or layer, there may be no intermediate elements or layers present. Other terms used to describe relationships between elements should be interpreted in a similar manner (e.g., “between” vs. “directly between,” “adjacent” vs. “directly adjacent,” etc.). As used herein, the term “and / or” includes any and all combinations of one or more of the associated listed items.

[0029] The terms “first,” “second,” “third,” etc., may be used herein to describe various elements, components, regions, layers, and / or sections. These elements, components, regions, layers, and / or sections should not be limited by these terms. These terms may be used only to distinguish one element, component, region, layer, or section from another. Unless the context clearly indicates otherwise, terms such as “first,” “second,” and other numerical terms do not imply order or sequence. Therefore, without departing from the teachings of the example configuration, the first element, component, region, layer, or section discussed below may be referred to as the second element, component, region, layer, or section.

[0030] In this application, including the following definitions, the term "module" may be replaced by the term "circuit". The term "module" may refer to, be a part of, or include: application-specific integrated circuits (ASICs); digital, analog, or mixed-signal analog / digital discrete circuits; digital, analog, or mixed-signal analog / digital integrated circuits; combinational logic circuits; field-programmable gate arrays (FPGAs); processors (shared, dedicated, or grouped) that execute code; memory (shared, dedicated, or grouped) that stores code executed by the processor; other suitable hardware components that provide the described functionality; or combinations of some or all of the foregoing, such as in a system-on-a-chip.

[0031] 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" covers a single processor that executes some or all of the code from multiple modules. The term "group processor" covers a processor that, in combination with additional processors, executes some or all of the code from one or more modules. The term "shared memory" covers a single memory that stores some or all of the code from multiple modules. The term "group memory" covers memory that, in combination with additional memory, 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" does not include transient electrical and electromagnetic signals propagating through the medium, and therefore can be considered tangible and non-transitory memory. Non-limiting examples of non-transitory memory include tangible computer-readable media, which include non-volatile memory, magnetic memory, and optical memory.

[0032] The apparatus and methods described in this application can be implemented, in part or in whole, by one or more computer programs executed by one or more processors. The computer program includes processor-executable instructions stored on at least one non-transitory tangible computer-readable medium. The computer program may also include and / or depend on stored data.

[0033] A software application (i.e., a software resource) can refer to computer software that enables a computing device to perform tasks. In some examples, a software application may be referred to as an "application," "app," or "program." Example applications include, but are not limited to, system diagnostic applications, system management applications, system maintenance applications, word processing applications, spreadsheet applications, messaging applications, media streaming applications, social networking applications, and game applications.

[0034] Non-transitory memory can be a physical device used to temporarily or permanently store programs (e.g., instruction sequences) or data (e.g., program state information) for use by a computing device. Non-transitory memory can be volatile and / or non-volatile addressable semiconductor memory. Examples of non-volatile memory include, but are not limited to, flash memory and read-only memory (ROM) / programmable read-only memory (PROM) / erasable programmable read-only memory (EPROM) / electrically erasable programmable read-only memory (EEPROM) (e.g., commonly used in firmware, such as bootloaders). Examples of volatile memory include, but are not limited to, random access memory (RAM), dynamic random access memory (DRAM), static random access memory (SRAM), phase-change memory (PCM), and magnetic disks or magnetic tapes.

[0035] These computer programs (also referred to as programs, software, software applications, or code) include machine instructions for a programmable processor and can be implemented using high-level procedural and / or object-oriented programming languages ​​and / or assembly / machine languages. As used herein, the terms “machine-readable medium” and “computer-readable medium” refer to any computer program product, non-transitory computer-readable medium, apparatus, and / or device (e.g., disk, optical disk, memory, programmable logic device (PLD)) used to provide machine instructions and / or data to a programmable processor, including machine-readable media that receive machine instructions as machine-readable signals. The term “machine-readable signal” refers to any signal used to provide machine instructions and / or data to a programmable processor.

[0036] Various implementations of the systems and techniques described herein can be implemented 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 implementations in one or more computer programs executable and / or interpretable on a programmable system, which includes at least one programmable processor, which may be dedicated or general-purpose, coupled to receive data and instructions from a storage system, at least one input device, and at least one output device, and to transmit data and instructions to the storage system, at least one input device, and at least one output device.

[0037] The processes and logic described in this specification can be executed by one or more programmable processors (also known as data processing hardware) that execute one or more computer programs to perform functions by manipulating input data and generating output. The processes and logic can also be executed by special-purpose logic circuitry, such as FPGAs (Field-Programmable Gate Arrays) or ASICs (Application-Specific Integrated Circuits). Processors suitable for executing computer programs include, for example, both general-purpose microprocessors and special-purpose microprocessors, as well as any one or more processors of any kind of digital computer. Typically, the processor receives instructions and data from read-only memory or random access memory, or both. The basic elements of a computer are a processor for executing instructions and one or more memory devices for storing instructions and data. Typically, a computer will also include one or more mass storage devices (e.g., magnetic disks, magneto-optical disks, or optical disks) for storing data, or operatively coupled to receive data from or transfer data to one or more mass storage devices, or both. However, a computer does not need to have such devices. Computer-readable media suitable for storing computer program instructions and data include all forms of non-volatile memory, media, and memory devices, including, for example, semiconductor memory devices such as EPROM, EEPROM, and flash memory devices; magnetic disks, such as internal hard disks or removable disks; magneto-optical disks; and CD-ROM and DVD-ROM disks. Processors and memory may be supplemented by or incorporated into dedicated logic circuitry.

[0038] To provide interaction with a user, one or more aspects of this disclosure can be implemented on a computer having a display device for displaying information to the user, such as a CRT (cathode ray tube), LCD (liquid crystal display) monitor, or touchscreen, and optionally a keyboard and pointing device, such as a mouse or trackball, through which the user can provide input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback, such as visual feedback, auditory feedback, or tactile feedback; and input from the user can be received in any form, including acoustic, voice, or tactile input. Furthermore, the computer can interact with the user by sending documents to and receiving documents from the device used by the user; for example, by sending a web page to a web browser on the user's client device in response to a request received from a web browser.

[0039] The development and validation of safe and reliable AVs rely heavily on rigorous testing across a wide range of driving scenarios. While real-world testing is essential, relying on it to cover a vast array of potential interactions, particularly rare but safety-critical "long-tail" events, is impractical and often dangerous. Therefore, high-fidelity simulation has become a fundamental tool for evaluating AV performance. Effective traffic simulators must generate scenarios that are both realistic and controllable. Realism ensures that the simulated behavior of surrounding agents (e.g., other vehicles, pedestrians) accurately reflects the complex, nuanced, and often unpredictable nature of real-world interactions. Controllability allows developers and testers to specifically create and analyze challenging situations, such as induced hazard or collision events, to systematically probe the limits of an AV's capabilities. However, existing traffic simulation methods struggle to adequately balance these two often competing objectives. Data-driven approaches that learn from large datasets of real-world driving can produce realistic, common behaviors but often fail to generate novel, safety-critical scenarios rarely encountered in the training data. Furthermore, when used in closed-loop environments where simulations evolve over time, these models can suffer from compounded errors, causing the simulation to drift into unrealistic states.

[0040] Conversely, rule-based methods offer precise control but often produce seemingly rote, rigid, and unrealistic behavior because they fail to capture the adaptive decision-making of human drivers. Recent deep generative models, including diffusion models, have shown promise but still face a fundamental challenge: the conflict between realism and controllability. The process of guiding a simulation toward a specific, user-defined outcome (e.g., a collision) typically requires generating proxy behaviors that are unlikely and significantly deviate from realistic patterns learned from data. This “gradient conflict” means that increasing controllability often comes at the direct cost of sacrificing realism, and vice versa. Therefore, a traffic scene generation system is needed that can resolve this conflict, enabling the creation of scenarios that are both highly realistic and precisely controllable, particularly for safety-critical events.

[0041] This system and method generate realistic and controllable closed-loop traffic scenarios, overcoming the aforementioned limitations of existing technologies. The system and method incorporate a diffusion model, called the Causal Combinatorial Diffusion Model (CCDiff), which addresses the inherent conflict between realism and controllability by identifying and utilizing the underlying causal structures of interactions within the traffic scenario. The method formulates the scenario generation task as a constrained optimization problem, aiming to maximize user-defined controllability objectives (e.g., inducing safety-critical events) while satisfying realistic constraints. At its core, the system employs a diffusion model to generate surrogate trajectories. The key innovation lies in how this generation process is guided.

[0042] First, the causal inference module analyzes the scene to automatically discover a decision causal graph (DCG), which maps causal relationships between agents. Based on this graph, agents are ranked according to their influence and interactivity, identifying a small subset of key agents crucial to achieving the desired outcome. Second, a novel causal combination guidance mechanism is used to guide the generation process of CCDiff. This guidance is structured and selective. Gradients associated with the controllability objective are applied only to the identified key agents. This centralized intervention effectively guides the scene toward the desired outcome. Simultaneously, the behavior of all other agents is guided by the identified causal structure, ensuring their actions remain consistent and realistic within the context of the scene. By decoupling the guidance mechanism in this way, the system avoids gradient conflicts that plague existing methods, thus achieving a better balance between realism and controllability.

[0043] refer to Figure 1 A system 100 for causal combination diffusion for closed-loop traffic generation is illustrated. System 100 includes a remote computing system 50 and an autonomous agent. While the autonomous agent is depicted as vehicle 10, the system and methods described herein are broadly applicable to other types of autonomous agents. Such agents can include, but are not limited to, autonomous mobile robots (AMRs) operating in warehouses, robotic manipulators performing tasks in dynamic environments, unmanned aerial vehicles (UAVs), or agricultural and construction equipment. Furthermore, the principles described can be applied to simulation systems for modeling agent behavior, such as in air traffic control systems or for pedestrian flow analysis. The remote computing system 50 can be a single computer, multiple computers, or a distributed system, such as a cloud computing environment, having data processing hardware 52 and memory 54. Memory 54 stores instructions that, when executed by the data processing hardware 52, configure the remote computing system 50 to operate as a causal combination diffusion model 200. The causal combination diffusion model 200 is configured to generate a scenario of future trajectories 234, which is then deployed to validate a driving model 120 executed by an onboard driver assistance system 12. Here, the future trajectory 234 can form the final traffic scenario 232 that satisfies the controllability objective of the causal combination diffusion model 200. As used herein, the controllability objective may include generating safety-critical events, such as, but not limited to, collisions, off-road incidents, near misses, or speeding incidents. Although described as a remote system, in some implementations, the functionality of the causal combination diffusion model 200 may be performed, in whole or in part, on computing resources located within the vehicle 10.

[0044] After the causal combination diffusion model 200 generates the future trajectory 234, the causal combination diffusion model 200 provides the future trajectory 234 for testing and validating inference tasks performed by the driving model 120. The future trajectory 234 can be deployed to a mobile platform, such as the vehicle 10 shown, for execution by the onboard controller 14. The disclosed method also extends beyond perception and control tasks. The vehicle controller 14 is part of an onboard control system, such as the driving 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 inference tasks involving processing real-time sensor data from the sensor system 20. The sensor system 20 may include various sensors, such as one or more cameras 22, radar sensors 24, or lidar sensors 26. The output of the inference task is provided to one or more functions of the driving assistance system 12, such as an adaptive cruise control system or an automatic emergency braking system.

[0045] refer to Figure 2 and Figure 3 The causal combination diffusion model 200 includes a scene encoder 210, a causal inferencer 300, a ranking module 220, and a diffusion model 230. The scene encoder 210 is configured to receive a history 208 of a previous traffic scene as input and perform structured scene encoding to encode the history 208 to generate a predicted action 212 for the history 208 as output. The causal combination diffusion model 300 is configured to receive an initial condition 202 of a traffic scene comprising multiple interacting agents 204 as input. Here, the initial condition 202 defines a state 206 for each of the multiple interacting agents 204. The causal combination diffusion model 300 is configured to identify a causal structure 332 among the multiple agents 204 based on the states 206 of the multiple agents 204.

[0046] like Figure 2 As shown, the causal inferencer 300 generates a causal structure 332 as output that defines the causal influences between agents 204 among a plurality of agents 204. In some embodiments, the causal structure 332 includes a decision causal graph (DCG) having multiple nodes representing agents 204 and edges representing causal dependencies of future actions. In some cases, the scene encoder 210 uses a factorization attention mechanism to generate the DCG of the causal structure 332. Here, causal connections can be identified based at least on one of the attention weights or kinematic factors among the plurality of agents 204. These kinematic factors may include collision time (TTC) values ​​between pairs of agents 204 among the plurality of agents 204.

[0047] refer to Figure 3The causal inferencer 300 may include a tokenizer 310, an attention layer 320, and a masking module 330 that collaborate to generate a causal structure 332 (e.g., a factorized DCG). Here, the causal inferencer 300 encodes the motion history 208 of different agents 204 in the initial condition 202 based on spatial attention, and then discovers the DCG based on the factorized attention mask and kinematic factors. Finally, the causal inferencer optimizes its controllability by masking non-critical agents 204 in order to guide the backsampling process of the diffusion model 230 in a structured manner. The tokenizer 310 includes a transformer-based structure configured to receive the initial condition 202 and the history 208 as input, and embed the history 208 of the agents 204 to generate an agent embedding 312. Here, the agent embedding 312 includes the history of each agent 204 relative to the history 208 of all other agents 204 in the initial condition 202. To facilitate relational reasoning, the absolute and relative features of agent 204 are combined, including the position, velocity, distance, and collision time (TTC) of each agent 204 relative to other agents 204.

[0048] Subsequently, attention layer 320 aggregates all temporal information from agent embedding 312 to generate attention output 322. To further discover useful spatial parent-child relationships, causal inferencer 300 applies two-step causal inference to identify DCG 332 in the spatial-temporal interactions of agent 204. First, hard constraints are set in adjacent perceptual fields by pruning unnecessary causal connections between agent state 206 and corresponding actions at time step t. Second, the first adjustable hard constraint is applied as a memory mask to the attention weights of agent 204 via masking module 330, as follows:

[0049] (1).

[0050] Here, M represents the memory mask extracted using relative TTC features. The surrounding agents 204 of each corresponding agent 204 in initial condition 202 are given a threshold C of DCG 332, as follows:

[0051] (2).

[0052] The masking module 330 can adjust the threshold C to control the sparsity of the final DCG 332s, so that the diffusion model 230 aggregates the map information C and the state of the causal parent agent 204 to obtain the final action of the future trajectory 234 of the agent 204.

[0053] Ranking module 220 is configured to receive a causal structure 332 comprising multiple agents 204 and their corresponding causal influences as input, and to rank the agents 204 to determine a subset of key agents 204K that are most influential with respect to the controllability objective of the causal combined diffusion model 200. In other words, ranking module 220 performs the top K guidances on the most influential agents 204 in the causal structure 332 to identify a subset of key agents 204K. In some cases, ranking module 220 performs graph-based analysis on the identified causal structure 332 to determine the degree of interaction of each of the multiple agents 204. Ranking module 220 may generate a subset of key agents 204K for diffusion model 230 as output.

[0054] Subsequently, the diffusion model 230 receives a causal structure 332, comprising a subset of agents 204 and key agents 204K, as input, and a predicted action 212 generated by the scene encoder 210 as input, and generates a future trajectory 232 for each agent 204 in the causal structure 332 as output. The diffusion model 230 can generate each future trajectory 232 using a backsampling process. In some cases, the diffusion model is a denoised diffusion probabilistic model (DDPM), where the backsampling process iteratively denoises the noise vector to produce each future trajectory 232 for agent 204. Notably, the causal combined diffusion model 200 is configured to guide the backsampling process of the diffusion model 230 by selectively applying the gradient of the controllability objective only to a subset of the determined key agents 204K. Here, the guidance for the remaining agents 204 is determined based on the causal structure 332 determined by the causal inferencer. By dividing the guidance of the backsampling process into two parts, the diffusion model 230 generates the final traffic scene 234 of all future trajectories 232 of the agent 204, which satisfies the controllability objective while maintaining realism.

[0055] The causal combination diffusion model 200 can decompose the controllability objective by formulating the closed-loop traffic simulation as a Markov decision process (MDP) problem, and utilize the diffusion model 230 ( Figure 2 Sequential modeling is performed to learn a controllable simulation policy. To utilize the causal structure between state 206, actions, and reward spaces, the causal combination diffusion model 200 is defined by a constrained factorization MDP and a decision causal graph (i.e., causal structure 332). A constrained factorization MDP (CFMDP) is a type of MDP where the state space S and reward function R are factored to utilize the structure of the problem. A CFMDP is defined by tuples: M... F= (S, A, P, R, C, s0). Here, S represents the factorized state space, which represents the trajectory space of each agent 204 at the current time step t. A represents the factorized action space, which includes interventions on the subsequent derived behaviors of each agent 204 in the scene. P represents the joint transition dynamics defined on the pair of states S and actions A, and defines the deterministic vehicle dynamics of each agent 204 in the setting. R represents the reward objective for collisions, off-road events, speeding, or other objectives, where each subset of R specifies the state factors affecting the reward. C represents the constraint function, which indicates the level of realism of the learned simulated policy relative to the generated trajectory of the dataset policy, where a lower constraint value implies higher realism. The initial state is represented by s0, which lies in the factorized state space s.

[0056] As described above, for each time step t, the causal structure 332 is defined as G, where G is true if and only if the future actions of a particular agent 204 are conditionally independent of the history 208 of agent 204. ij =0. In G 2,3 When = 1, the causal structure 332 includes causal edges for that particular agent 204. As described above, the causal combination diffusion model 200 defines a set of policies, where the causal parent of each agent 204 in the causal structure 332 is used to make decisions when identifying the causal structure 332. Given CFMDP, and utilizing known vehicle dynamics, the causal combination diffusion model 200 factorizes the objective of generating the optimal closed-loop scenario as follows:

[0057]

[0058] The first term corresponds to controllability (i.e., the probability of optimality specified by a reward objective defined by some users), and the second term corresponds to realism (i.e., the probability of the behavior generated in future trajectory 232). Therefore, the scoring function for the maximum likelihood objective can be expressed as follows:

[0059] (4).

[0060] Unlike the normal situation where optimization and imitation largely conform to the rules and rewards, safety-critical guidance R (3) Gradient conflicts may occur. To address these gradient conflict issues, the causal combinatorial diffusion model 200 prioritizes controlling the index of agent 204 and maximizes the reward while maintaining high likelihood of the learned policy (i.e., a low truth gap between the learned policy and the behavioral policy). Here, the diffusion model 230 can use Lagrange multipliers and structured projective gradient descent to solve the constrained optimization problem with the following maximum likelihood estimation problem:

[0061] (5),

[0062] Satisfying |G|≤ This can be changed The level of constraint is used to control the level of authenticity.

[0063] Figure 4 A flowchart illustrating an exemplary arrangement of the operation of a method 400 for causal combination diffusion in closed-loop traffic generation is provided. See also... Figure 1-3 Description method 400. Data processing hardware (e.g., Figure 1 Data processing hardware 52) can execute data stored in memory hardware (e.g., Figure 1 An example arrangement of instructions on memory hardware 54 to perform the operations of method 400. At operation 402, method 400 includes receiving initial conditions 202 of a traffic scenario including a plurality of interacting agents 204, the initial conditions 202 defining states 206 of the plurality of agents 204. At operation 404, method 400 includes identifying a causal structure 332 among the plurality of agents 204 based on the states 206 of the plurality of agents 204. Here, the causal structure 332 defines the causal effects among the agents 204.

[0064] At operation 406, method 400 further includes ranking the multiple agents 204 based on the identified causal structure 332 to determine a subset of key agents 204K that are most influential on the controllability objective. For each of the multiple agents 204, method 400 further includes generating a future trajectory 232 using a backsampling process of diffusion model 230 at operation 408. At operation 410, method 400 further includes guiding the backsampling process by selectively applying the gradient of the controllability objective only to the determined subset of key agents 204K. Here, guidance for the remaining agents 204 among the multiple agents 204 is determined based on the identified causal structure 332, thereby generating a final traffic scene 234 that satisfies the controllability objective while maintaining realism.

[0065] Figure 5 A flowchart illustrating an exemplary arrangement of the operation of a method 500 for causal combination diffusion in closed-loop traffic generation is provided. See also... Figure 1-3 To describe method 500. Data processing hardware (e.g., Figure 1 Data processing hardware 52) can execute data stored in memory hardware (e.g., Figure 1An example arrangement of instructions on memory hardware 54 to perform the operations of method 500. At operation 502, method 500 includes receiving initial conditions 202 of a traffic scene including a plurality of interacting agents 204, the initial conditions 202 defining states 206 of the plurality of agents 204. At operation 504, method 500 includes identifying a causal structure 332 among the plurality of agents 204 based on the states 206 of the plurality of agents 204, where the causal structure 332 defines causal influences among the agents 204.

[0066] At operation 506, method 500 further includes ranking multiple agents 504 based on the identified causal structure 332 to determine a subset of key agents 204K that are most influential regarding the controllability objective. For each of the multiple agents 204, method 500 further includes, at operation 508, generating a future trajectory 232 using a backsampling process of the diffusion model 230. At operation 510, method 500 further includes guiding the backsampling process by selectively applying the gradient of the controllability objective only to the determined subset of key agents 204K.

[0067] Many embodiments have been described. However, it should be understood that various modifications can be made without departing from the spirit and scope of this disclosure. Therefore, other embodiments are within the scope of the appended claims.

[0068] The foregoing description has been provided for purposes of illustration and description. It is not intended to be exhaustive or limiting of this disclosure. Elements or features of a particular configuration are generally not limited to that particular configuration, but are interchangeable where applicable and can be used in selected configurations, even if not specifically shown or described. They can also be varied in many ways. Such variations should not be considered as departing from this disclosure, and all such modifications are intended to be included within the scope of this disclosure.

Claims

1. A computer-implemented method, when executed on data processing hardware, causes the data processing hardware to perform operations, the operations including: Receive initial conditions for a traffic scenario that include multiple interactive agents, wherein the initial conditions define the states of the multiple agents; The causal structure among the multiple agents is identified based on their states, and the causal structure defines the causal influence among the agents. The multiple agents are ranked based on the identified causal structure to determine a subset of key agents that have the greatest impact on the controllability objective; For each of the multiple agents, the future trajectory is generated using the backsampling process of the diffusion model; and The backsampling process is guided by selectively applying the gradient of the controllability objective only to a subset of the identified key agents, while determining the guidance for the remaining agents among the plurality of agents based on the identified causal structure, thereby generating a final traffic scene that satisfies the controllability objective while maintaining realism.

2. The method according to claim 1, wherein, Identifying the causal structure includes generating a decision causal graph (DCG), where nodes represent agents and edges represent causal dependencies of future actions.

3. The method according to claim 2, wherein, The DCG is generated using a scene encoder with a factorization attention mechanism, wherein causal connections are identified based on at least one of the attention weights or kinematic factors among the plurality of agents.

4. The method according to claim 3, wherein, The kinematic factors include the collision time (TTC) values ​​between pairs of agents among the plurality of agents.

5. The method according to claim 1, wherein, Ranking the plurality of agents includes performing graph-based analysis on the identified causal structure to determine the degree of interactivity of each of the plurality of agents.

6. The method according to claim 1, wherein, The backsampling process also includes the application of a classifier-free guiding component, which comprises a weighted combination of an unconditional distribution based on the agent's own history and an intervention distribution based on the agent's causal parents as defined by the causal structure.

7. The method according to claim 1, wherein, The controllability objective is associated with the generation of safety-critical events.

8. The method according to claim 7, wherein, The safety-critical events include one of the following: a collision between at least two agents, an off-road incident involving at least one agent, or a near-miss incident.

9. The method according to claim 1, wherein, The method is formulated as a constrained optimization problem in a constrained factor Markov decision process (CFMDP), where the controllability objective is maximized under realistic constraints.

10. The method according to claim 1, wherein, The diffusion model is a denoised diffusion probability model (DDPM), and the backsampling process iteratively denoises the noise vector to generate each of the future trajectories.

Images