Knowledge Distillation-Enhanced Behavior Transformer for Decision-Making of Autonomous Driving

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Autonomous driving relies on behavior decision-making, bridging perception and control.
• Imitation Learning (IL) and Reinforcement Learning (RL) are key approaches, but RL faces challenges in complex environments due to limited reasoning and sample efficiency.

Purpose of the Study:

• To propose an innovative Knowledge Distillation-Enhanced Behavior Transformer (KD-BeT) framework.
• To leverage Transformer's contextual reasoning for sequential decision-making in autonomous driving.
• To improve RL's training efficiency and performance in complex driving scenarios.

Main Methods:

• Introduced a Behavior Transformer as the policy network in RL, utilizing observation-action history.
• Employed a teacher-student paradigm: a teacher model trained via IL, followed by knowledge distillation to accelerate RL.
• Applied the KD-BeT framework to autonomous driving behavior decision-making.

Main Results:

• KD-BeT demonstrated fast convergence and high asymptotic performance during training.
• Outperformed state-of-the-art methods in CARLA NoCrash benchmark tests for traffic efficiency and driving safety.
• Validated the effectiveness of knowledge distillation in enhancing RL for autonomous driving.

Conclusions:

• The KD-BeT framework offers a novel and effective solution for autonomous driving behavior decision-making.
• Successfully combined Transformer architecture with knowledge distillation to overcome RL limitations.
• Achieved superior performance in traffic efficiency and driving safety, paving the way for real-world applications.