RAKD: Risk-Aware Knowledge Distillation for Autonomous Driving in Adversarial Scenarios

Aggressive driving behaviors of surrounding road users pose substantial threats to the safety of autonomous vehicles. However, current imitation learning methods struggle to learn a risk-aware driving policy, as they fail to learn the intrinsic decision-making principles and safety-relevant features from high-dimensional records. To address this issue, we propose a robust risk-aware knowledge distillation (RAKD) framework. RAKD employs a privileged teacher based on reinforcement learning to supervise a vision-based student via policy and feature distillation. The privileged teacher leverages two novel safetycritical representations in bird's-eye view and risk-aware reward shaping to enhance risk-sensitive learning, establishing superior imitation benchmarks through on-policy learning. Building on these benchmarks, we integrate a teachable module in the student network, aligning with the teacher's features through multiscale attention-based feature distillation, focusing on collision risk-related information. Experimental results demonstrate that our RAKD-trained student achieves superior driving score and collision avoidance in diverse adversarial scenarios compared to baselines. Our repository is publicly available at https://github. com/quliang93/RAKD.