HAD-Gen: Human-like and diverse driving behavior modeling for controllable scenario generation

Area of Science:

• Robotics
• Artificial Intelligence
• Transportation Engineering

Background:

• Simulation-based testing is crucial for autonomous vehicle (AV) verification and validation.
• Current driver models (deterministic, imitation learning) fail to capture human driving variability.
• Realistic traffic scenario generation is needed to address these limitations.

Purpose of the Study:

• To propose HAD-Gen, a general framework for generating realistic traffic scenarios with diverse human-like driving behaviors.
• To enhance the simulation-based testing of autonomous vehicles.
• To improve the generalization capability of driving policies in unseen scenarios.

Main Methods:

• Clustering vehicle trajectory data into distinct driving styles based on safety features.
• Applying maximum entropy inverse reinforcement learning to learn driving-style-specific reward functions.
• Integrating offline reinforcement learning pre-training and multi-agent reinforcement learning for robust driving policies.

Main Results:

• HAD-Gen successfully generates diverse, human-like driving behaviors in highway scenarios.
• The framework demonstrates strong generalization capabilities in new, unseen driving scenarios.
• Achieved a 90.96% goal-reaching rate, 2.08% off-road rate, and 6.91% collision rate, outperforming prior methods by over 20% in goal-reaching.

Conclusions:

• The proposed HAD-Gen framework effectively simulates human-like driving behaviors for realistic AV testing.
• HAD-Gen significantly improves the performance and generalization of autonomous driving policies.
• This approach offers a robust solution for generating complex traffic scenarios essential for AV development.