Agent autonomous confrontation behavior-based automatic driving edge scene generation method and system

Abstract

The application discloses an automatic driving edge scene generation method and system based on autonomous confrontation behavior of an agent, and belongs to the technical field of automatic driving. The method comprises the following steps: generating a scene joint probability distribution according to the states of traffic participants in a traffic scene and a to-be-tested automatic driving system, and sampling the scene joint probability distribution to obtain an initial parameter distribution of a generative adversarial network; designing a reward function, wherein a reward value of the reward function is positively correlated with a collision risk of the scene; performing parameter distribution optimization on the generative adversarial network based on the reward function, and generating a test scene of the to-be-tested automatic driving system based on the optimized generative adversarial network. The application can autonomously generate an edge scene with high risk and high test value, and can perform quantitative evaluation on an automatic driving system.

Description

Technical Field

[0001] This invention belongs to the field of autonomous driving technology, specifically relating to a method and system for generating autonomous driving edge scenes based on the autonomous adversarial behavior of intelligent agents. Background Technology

[0002] With the development of autonomous driving technology, its large-scale deployment in the real world faces challenges in safety assessment. Currently, the industry generally adopts the method of reproducing real-world scene logs to obtain edge scenes, but this method, which relies on massive data playback, has significant drawbacks in practical applications.

[0003] First, real-world scenario data is extremely vast and complex, involving ever-changing road conditions, weather conditions, and traffic participant behavior patterns. This requires manufacturers to spend a lot of manpower and resources on data collection and storage, increasing costs and extending the development cycle. Since safety-critical scenarios such as traffic accidents are extremely rare in daily driving, relying solely on real-world data is insufficient to fully cover all potential dangerous scenarios. As a result, autonomous driving systems often perform poorly when faced with unprecedented extreme conditions. Furthermore, the lack of uniformity in data standards among manufacturers makes system performance evaluations lack fairness and comparability.

[0004] Secondly, scene generation technologies that have emerged to address the aforementioned problem of insufficient data coverage, whether data-driven or adversarial, still have technical limitations. Scenes generated by data-driven methods often lack diversity and representativeness, making it difficult to cover complex and dangerous situations. While existing adversarial generation methods attempt to discover system vulnerabilities by introducing adversarial agents to compete with autonomous driving systems, they mostly focus on adversarial challenges against static object shapes in perception algorithms or simple trajectory perturbations in path planning. These studies often only focus on how to create collision results, and the generated attack trajectories often violate physical laws or human behavior habits, causing traffic participants' behavior to deviate significantly from the logic and characteristics of the real world. Such "collision for the sake of collision" scenarios, lacking realism and interpretability, fail to truly reflect the system's ability to cope in complex human traffic environments.

[0005] Finally, current simulation testing platforms lack a systematic quantitative evaluation index system. Most existing assessments focus on the single outcome of whether a collision occurs. For key dimensions such as comfort, risk level, and driving efficiency of autonomous driving systems in the process of risk avoidance, there is a lack of unified quantitative calculation standards. This makes it difficult for researchers to accurately measure and compare the performance of different autonomous driving algorithms in edge scenarios.

[0006] Therefore, there is an urgent need for a method to generate edge scenarios that are both high-risk and challenging and conform to the behavioral logic of real-world intelligent agents, and to combine it with a multi-dimensional quantitative evaluation system to achieve an efficient, realistic and quantifiable assessment of the safety of autonomous driving systems. Summary of the Invention

[0007] This invention proposes a method and system for generating edge scenes for autonomous driving based on the autonomous adversarial behavior of intelligent agents. It utilizes probabilistic graphical models and reinforcement learning adversarial strategies to autonomously generate edge scenes with high risk and high testing value, and to quantitatively evaluate autonomous driving systems.

[0008] To achieve the above objectives, the technical solution of the present invention includes the following:

[0009] A method for generating edge scenes for autonomous driving based on the autonomous adversarial behavior of intelligent agents, the method comprising: Based on the states of traffic participants and the autonomous driving system under test in the traffic scenario, a joint probability distribution of the scenario is generated, and the initial parameter distribution of the adversarial generative model is obtained by sampling from the joint probability distribution of the scenario; wherein, the adversarial generative model is used to generate the trajectories of traffic participants; Design a reward function whose reward value is positively correlated with the collision risk of the scene; Based on the reward function, the parameter distribution of the adversarial generative model is optimized, and based on the optimized adversarial generative model, a test scenario for the autonomous driving system to be tested is generated.

[0010] Furthermore, based on the states of traffic participants and the autonomous driving system under test in the traffic scenario, a joint probability distribution for the scenario is generated, including: Traffic participants are decomposed into several elements, including starting location, final destination, and individual characteristics; Based on the logical relationships between elements and between elements and the state of the autonomous driving system under test, a joint probability distribution of the scenario is generated.

[0011] Furthermore, the reward value ;in, For risk measurement, For the autonomous driving system to be tested, For traffic participants, For collision rewards, As a reasonable penalty item, The threshold value is used.

[0012] Furthermore, based on this reward function, the parameter distribution of the adversarial generative model is optimized, including: Generate parameters for the traffic participant agent based on the current parameter distribution of the adversarial generative model; Deploy the parameters of traffic participant intelligent agents into the simulation platform; The autonomous driving system under test is connected to the simulation platform and interacts with the intelligent agents of traffic participants to obtain the interaction results; Calculate the reward value based on the interaction results and the reward function; The current parameter distribution of the adversarial generative model is updated based on the reward value; Based on the updated parameter distribution, the step of generating parameters for the traffic participant agent according to the current parameter distribution of the adversarial generative model is re-executed.

[0013] Furthermore, after generating test scenarios for the autonomous driving system to be tested based on the optimized adversarial generative model, the method further includes: Based on the vehicle's acceleration comfort and yaw rate comfort, the comfort evaluation index of the autonomous driving system under test is obtained. Based on the minimum distance between the vehicle and the nearest obstacle, calculate the risk assessment index of the autonomous driving system under test; The driving speed evaluation index is obtained based on the deviation between the vehicle's average speed and the target speed; The evaluation results of the autonomous driving system under test are obtained by combining the comfort evaluation index, the risk level evaluation index, and the driving speed evaluation index.

[0014] An autonomous driving edge scene generation system based on agent-driven autonomous adversarial behavior, the system comprising: The model building module is used to generate a joint probability distribution of the scene based on the state of traffic participants and the autonomous driving system under test in the traffic scene, and to sample from the joint probability distribution of the scene to obtain the initial parameter distribution of the adversarial generative model; wherein, the adversarial generative model is used to generate the trajectory of traffic participants; The reward design module is used to design a reward function whose reward value is positively correlated with the collision risk of the scene. The scenario generation module is used to optimize the parameter distribution of the adversarial generative model based on the reward function, and generate test scenarios for the autonomous driving system to be tested based on the optimized adversarial generative model.

[0015] An electronic device includes: a processor and a memory storing computer program instructions; the processor, when executing the computer program instructions, implements the autonomous driving edge scene generation method based on the autonomous adversarial behavior of intelligent agents as described above.

[0016] A computer-readable storage medium storing computer program instructions, which, when executed by a processor, implement the autonomous driving edge scene generation method based on the autonomous adversarial behavior of intelligent agents as described above.

[0017] A computer program product, characterized in that, when the computer program product is run on a computer device, it causes the computer device to execute the autonomous driving edge scene generation method based on the autonomous adversarial behavior of intelligent agents as described above.

[0018] Compared with existing technologies, this invention, through an adversarial game mechanism, can automatically search for and generate high-risk edge scenarios that are rare in reality but crucial in testing, significantly reducing the reliance on massive amounts of real-world data; it introduces a probabilistic graphical model and a rationality penalty term to ensure that the behavior of the generated agent conforms to real-world logic and behavioral characteristics, avoiding the problem of chaotic trajectories in traditional adversarial generation; and it proposes multi-dimensional quantitative indicators covering comfort, risk, and efficiency, solving the problem of the lack of a unified evaluation standard in current simulation testing, and enabling a more comprehensive measurement of the merits of different autonomous driving systems. Attached Figure Description

[0019] Figure 1 A flowchart of an autonomous driving edge scene generation method based on agent autonomous adversarial behavior.

[0020] Figure 2 Schematic diagram of the adversarial generative framework.

[0021] Figure 3 Example of generated edge scene.

[0022] Figure 4 A radar chart showing the quantitative scores of different autonomous driving systems across three dimensions: comfort, risk level, and driving speed.

[0023] Figure 5 Block diagram of an autonomous driving edge scene generation system based on the autonomous adversarial behavior of intelligent agents.

[0024] Figure 6 Block diagram of an electronic device. Detailed Implementation

[0025] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be noted that the embodiments described below are intended to facilitate the understanding of the present invention and do not limit it in any way.

[0026] This invention constructs an adversarial game between an intelligent agent and an autonomous driving system, uses a probabilistic graphical model to describe the scenario, and combines a reward function that includes risk rewards and reasonableness penalties to autonomously search for and generate high-risk test scenarios that can trigger the failure of the autonomous driving system. Simultaneously, it establishes a quantitative evaluation index system that includes comfort, risk level, and driving speed.

[0027] like Figure 1 As shown, the autonomous driving edge scene generation method based on the autonomous adversarial behavior of intelligent agents of the present invention includes the following steps.

[0028] Step S1: Generate a joint probability distribution of the scene based on the states of traffic participants and the autonomous driving system under test in the traffic scene, and sample from the joint probability distribution of the scene to obtain the initial parameter distribution of the adversarial generative model.

[0029] This embodiment aims to address the challenges of scene description and diverse generation. First, traffic participants in a traffic scene are decomposed into modules such as starting position, final destination, and individual characteristics. A directed graph is used to simplify the representation, and conditional probabilities are used to describe node dependencies. For example, in a forked intersection scene, a potentially risky pedestrian scene can be represented by four nodes: the pedestrian's starting x-coordinate... Initial ordinate Ultimate goal and individual characteristics , The number of individual features. Joint probability of the scene. Represented as: in, This represents the state of the autonomous driving system under test. Diverse base scenarios can be generated by sampling from this joint distribution.

[0030] Furthermore, to generate edge scenarios with high testing value, this embodiment constructs an adversarial generative model. For example... Figure 2 As shown, it searches for high-risk scenarios by adjusting parameters, and the entire process is optimized using the policy gradient method.

[0031] Step S2: Design a reward function whose reward value is positively correlated with the collision risk of the scene.

[0032] The core of adversarial generative models is the reward function. The reward function formula in this embodiment is as follows: in: As a risk metric, the distance between autonomous vehicles and traffic participants is calculated. This indicates that the closer the distance, the higher the reward; therefore, the formula takes a negative sign. Wherein, The location of the autonomous vehicle. Location of traffic participants; As a collision reward, an additional reward is given if a collision occurs during the simulation (collision=True). Otherwise, it is 0; This is a reasonable penalty item used to control the deviation distance of traffic participants. If the deviation amount... Less than the threshold The punishment is Otherwise, it is 0. This represents the i-th new location point generated by the traffic participant, thereby avoiding the generation of unreasonable teleportation or scenarios that violate the laws of physics.

[0033] Step S3: Optimize the parameter distribution of the adversarial generative model based on the reward function, and generate a test scenario for the autonomous driving system to be tested based on the optimized generative model.

[0034] Sampling: The sampler generates parameters for intelligent agents such as pedestrians or vehicles based on the current parameter distribution.

[0035] Deployment and Communication: The generated agent parameters are deployed to simulation platforms such as CARLA via the Cyber ​​RT communication module.

[0036] Interactive Game Theory: The autonomous driving system under test (such as Apollo or Autoware) connects to the simulation platform and interacts with the generated intelligent agent.

[0037] Feedback Update: The reward function calculation module calculates the reward value based on the interaction results, such as whether there is a collision and the distance, and updates the parameters of the generated model to make it more likely to generate high-risk scenarios in the next round.

[0038] Therefore, it is possible to generate, such as Figure 3 The following are edge scene scenarios. In (a), a pedestrian stops because he sees a purple vehicle speeding towards him, which causes a black vehicle turning the corner, which the pedestrian did not notice, to brake suddenly and scrape against the purple vehicle; in (b), a pedestrian is crossing the road at a bend, and the turning vehicle does not react in time, resulting in a collision.

[0039] In one embodiment of the present invention, the method further includes: step S4: quantitatively evaluating the autonomous driving system under test based on the test scenario.

[0040] To evaluate the performance of the autonomous driving system in the generated scenario, this embodiment defines three evaluation metrics: Comfort ): Evaluate the smoothness of operation. Combine with acceleration comfort. ( (Integral of axial acceleration) and yaw rate comfort The formula is: , , Indicates the weight.

[0041] Risk level ( Assess potential hazards. Based on the minimum distance between the vehicle and the nearest obstacle. The comprehensive risk index is the mean of the reciprocals of the minimum distances across multiple scenarios: , Represents the minimum distance in the i-th scene. N represents the total number of scenes.

[0042] driving speed ( ): Evaluate efficiency. Combine with average speed Deviation from target speed The formula is: , This represents the average speed weighting coefficient. This represents the target speed deviation weighting coefficient.

[0043] like Figure 4 As shown in the experiment, the scenarios generated using this method (such as pedestrians suddenly peeking out or crossing the road at a bend) can effectively distinguish the performance of different systems. For example, the test found that the Apollo system is better than other systems in terms of driving efficiency, but its behavior is more aggressive; while the Autoware system is slightly safer, but tends to adopt inefficient strategies.

[0044] Based on the same concept, this invention also discloses an autonomous driving edge scene generation system based on the autonomous adversarial behavior of intelligent agents, such as... Figure 5 As shown, the system includes: The model building module is used to generate a joint probability distribution of a traffic scenario based on the states of traffic participants and the autonomous driving system under test, and to sample from the joint probability distribution of the scenario to obtain the initial parameter distribution of the adversarial generative model; wherein, the adversarial generative model is used to generate the parameters of traffic participants; The reward design module is used to design a reward function whose reward value is positively correlated with the collision risk of the scene. The scenario generation module is used to optimize the parameter distribution of the adversarial generative model based on the reward function, and generate test scenarios for the autonomous driving system to be tested based on the optimized adversarial generative model.

[0045] Based on the same concept, this invention also discloses a computer device, which may be a terminal, a laptop computer, a desktop computer, a server, a computer cluster, or other types of computer devices. For example... Figure 6 As shown, the computer device may include at least one processor and memory. The processor can execute instructions stored in the memory. The processor is communicatively connected to the memory via a data bus. In addition to the memory, the processor can also be communicatively connected to input devices, output devices, and communication devices via the data bus.

[0046] The processor can be any conventional processor. Processors may include central processing units (CPUs), graphics processing units (GPUs), field-programmable gate arrays (FPGAs), systems on chips (SoCs), application-specific integrated circuits (ASICs), or combinations thereof.

[0047] Memory can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk.

[0048] In this embodiment of the invention, an executable instruction is stored in a memory. The processor can read the executable instruction from the memory and execute the instruction to implement all or part of the steps of the method of the invention.

[0049] Based on the same concept, the present invention also discloses a computer-readable storage medium including a computer program product or storing the computer program product. The computer product includes computer program instructions that can be executed by a processor to perform all or part of the steps described in the exemplary embodiments above.

[0050] Computer program products can be written in any combination of one or more programming languages ​​to perform the operations of the embodiments of this application. These programming languages ​​include object-oriented programming languages ​​such as Java and C++, as well as conventional procedural programming languages ​​such as C or similar languages, and scripting languages ​​(e.g., Python). The program code can be executed entirely on the user's computing device, partially on the user's device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0051] Computer-readable storage media can take the form of any combination of one or more readable media. A readable medium can be a readable signal medium or a readable storage medium. A readable storage medium can be, for example, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of readable storage media include: static random access memory (SRAM) having one or more electrically connected wires; electrically erasable programmable read-only memory (EEPROM); erasable programmable read-only memory (EPROM); programmable read-only memory (PROM); read-only memory (ROM); magnetic storage; flash memory; magnetic disk or optical disk; or any suitable combination thereof.

[0052] The above embodiments are merely illustrative of the technical solutions of the present invention and are not intended to limit it. Those skilled in the art can modify or make equivalent substitutions to the above technical solutions based on the concept of the present invention, and such modifications or equivalent substitutions should all be covered within the protection scope of the present invention. The protection scope of the present invention is defined by the claims.

Claims

1. A method for generating edge scenes for autonomous driving based on the autonomous adversarial behavior of intelligent agents, characterized in that, The method includes: Based on the states of traffic participants and the autonomous driving system under test in the traffic scenario, a joint probability distribution of the scenario is generated, and the initial parameter distribution of the adversarial generative model is obtained by sampling from the joint probability distribution of the scenario; wherein, the adversarial generative model is used to generate the trajectories of traffic participants; Design a reward function whose reward value is positively correlated with the collision risk of the scene; Based on the reward function, the parameter distribution of the adversarial generative model is optimized, and based on the optimized generative model, a test scenario for the autonomous driving system to be tested is generated.

2. The method according to claim 1, characterized in that, Based on the states of traffic participants and the autonomous driving system under test in the traffic scenario, a joint probability distribution for the scenario is generated, including: Traffic participants are decomposed into several elements, including starting location, final destination, and individual characteristics; Based on the logical relationships between elements and between elements and the state of the autonomous driving system under test, a joint probability distribution of the scenario is generated.

3. The method according to claim 1, characterized in that, The reward value ;in, For risk measurement, For the autonomous driving system to be tested, For traffic participants, For collision rewards, As a reasonable penalty item, The threshold value is used.

4. The method according to claim 1, characterized in that, Based on this reward function, the parameter distribution of the adversarial generative model is optimized, including: Generate parameters for the traffic participant agent based on the current parameter distribution of the adversarial generative model; Deploy the parameters of traffic participant intelligent agents into the simulation platform; The autonomous driving system under test is connected to the simulation platform and interacts with the intelligent agents of traffic participants to obtain the interaction results; Calculate the reward value based on the interaction results and the reward function; The current parameter distribution of the adversarial generative model is updated based on the reward value; Based on the updated parameter distribution, the step of generating parameters for the traffic participant agent according to the current parameter distribution of the adversarial generative model is re-executed.

5. The method according to any one of claims 1 to 4, characterized in that, After generating test scenarios for the autonomous driving system to be tested based on the optimized adversarial generative model, the method further includes: Based on the vehicle's acceleration comfort and yaw rate comfort, the comfort evaluation index of the autonomous driving system under test is obtained. Based on the minimum distance between the vehicle and the nearest obstacle, calculate the risk assessment index of the autonomous driving system under test; The driving speed evaluation index is obtained based on the deviation between the vehicle's average speed and the target speed; The evaluation results of the autonomous driving system under test are obtained by combining the comfort evaluation index, the risk level evaluation index, and the driving speed evaluation index.

6. An autonomous driving edge scene generation system based on the autonomous adversarial behavior of intelligent agents, characterized in that, The system includes: The model building module is used to generate a joint probability distribution of the scene based on the state of traffic participants and the autonomous driving system under test in the traffic scene, and to sample from the joint probability distribution of the scene to obtain the initial parameter distribution of the adversarial generative model; wherein, the adversarial generative model is used to generate the trajectory of traffic participants; The reward design module is used to design a reward function whose reward value is positively correlated with the collision risk of the scene. The scenario generation module is used to optimize the parameter distribution of the adversarial generative model based on the reward function, and generate test scenarios for the autonomous driving system to be tested based on the optimized adversarial generative model.

7. An electronic device, characterized in that, The electronic device includes: a processor and a memory storing computer program instructions; when the processor executes the computer program instructions, it implements the autonomous driving edge scene generation method based on the autonomous adversarial behavior of intelligent agents as described in any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer program instructions, which, when executed by a processor, implement the autonomous driving edge scene generation method based on the autonomous adversarial behavior of intelligent agents as described in any one of claims 1-5.

9. A computer program product, characterized in that, When the computer program product is run on a computer device, the computer device performs the autonomous driving edge scene generation method based on the autonomous adversarial behavior of intelligent agents as described in any one of claims 1-5.

Images