Human-guided continual learning for multifaceted improvement of self-driving vehicles

Today's self-driving vehicles have achieved impressive driving capabilities, nonetheless, safety concerns arising from ambiguous traffic laws, rare long-tail events, etc., still pose a significant challenge to their practical deployment. Therefore, human drivers are necessitated to take over in certain cases. Here, we present a human-guided continual learning method that leverages these human guidance data to continually improve self-driving performance, thereby enabling better handling of similar cases in the future. Our technique facilitates performance improvement by merely using the new data collected during driving, without requiring lengthy re-training from scratch. We evaluate the proposed technology through both simulations and real-world experiments, showing that it enables continual improvement by incrementally acquiring small amounts of human guidance. After each learning stage, the updated policy matches or outperforms the previous self-driving policy in terms of social compliance, rare case handling, etc. These findings highlight the potential of this technology to continually improve self-driving vehicles across multiple dimensions and to support broader human-in-the-loop autonomous systems.