Dense reinforcement learning for safety validation of autonomous vehicles

Area of Science:

• Artificial Intelligence
• Autonomous Systems
• Machine Learning

Background:

• Autonomous vehicle safety validation is costly and time-consuming due to rare critical events.
• Current methods struggle to efficiently test safety-critical scenarios.

Purpose of the Study:

• To develop an intelligent testing environment for accelerated autonomous vehicle safety validation.
• To train AI agents to identify and execute adversarial maneuvers.

Main Methods:

• Developed a dense deep-reinforcement learning (D2RL) approach.
• Trained background agents using naturalistic driving data to learn adversarial maneuvers.
• Utilized an augmented-reality environment combining simulated and real-world elements for testing.

Main Results:

• D2RL-trained agents accelerated the safety evaluation process by 10^3 to 10^5 times.
• The approach maintained unbiasedness in safety performance validation.
• Demonstrated effectiveness in both highway and urban test tracks.

Conclusions:

• D2RL offers a highly efficient method for autonomous vehicle safety validation.
• This approach significantly reduces economic and time costs for testing.
• D2RL is applicable to accelerating testing and training for other safety-critical autonomous systems.