A dynamic test scenario generation method for autonomous vehicles based on conditional generative adversarial...

Area of Science:

• Robotics
• Artificial Intelligence
• Transportation Engineering

Background:

• Current autonomous vehicle evaluation methods are often static, failing to capture crucial time-sequential interactions with environmental vehicles.
• Predefined trajectories in existing models do not reflect the dynamic and human-like behaviors of surrounding traffic.

Purpose of the Study:

• To propose a novel dynamic test scenario generation method for comprehensive autonomous vehicle evaluation.
• To address the limitations of static evaluation approaches by incorporating realistic vehicle interactions.

Main Methods:

• Environmental vehicles are modeled as agents exhibiting human-like behavior.
• A conditional generative adversarial imitation learning (CGAIL) model is employed, integrating clustered traffic scenario features.
• The method simulates the interaction process between the autonomous vehicle and environmental agents.

Main Results:

• The proposed method effectively generates dynamic test scenarios, particularly for complex interactions like lane changes.
• Validated results demonstrate enhanced testing of autonomous vehicles' adaptability to dynamic environments.
• The approach successfully identifies weaknesses in tested autonomous vehicles.

Conclusions:

• Dynamic scenario generation is crucial for robust autonomous vehicle testing.
• The proposed method, leveraging CGAIL and human-like agent modeling, offers a significant advancement in evaluation techniques.
• This approach provides a pathway to more thorough and insightful testing of autonomous driving systems.