Reproduce “DriveAdapter: Breaking the Coupling Barrier of Perception and Planning in End-to-End Autonomous Driving” on OpenResearch

About this paper

Xiaosong Jia, Yulu Gao, Li Chen, Junchi Yan, Patrick Langechuan Liu, Hongyang Li

DriveAdapter introduces a novel end-to-end autonomous driving framework that directly leverages a frozen, reinforcement learning-trained teacher for planning, enabling a student model to focus purely on perception learning. This approach effectively bridges the distribution gap between perception outputs and teacher inputs, mitigating causal confusion and achieving state-of-the-art performance on CARLA benchmarks, including a Driving Score of 71.9 on Town05 Long with extended data.

Key insights

• Directly feeding imperfect perception outputs to a frozen teacher model yields very poor performance, underscoring the necessity of explicitly bridging the 'distribution gap' through adaptive mechanisms.
• The combination of layer-wise feature alignment and action guidance is crucial for effectively transferring the teacher's robust planning knowledge and incorporating hand-crafted safety rules, mitigating causal confusion and enhancing robustness.
• Freezing the pre-trained teacher model is essential to preserve its extensive reinforcement learning-derived driving knowledge, with the adapter mechanism providing the interface without catastrophic forgetting.

Summary by alphaXiv.