Dynamic Scenario Transfer for Motion Planning in Autonomous Driving: A Successor Representation Approach

Adaptive motion planning systems are essential for autonomous vehicles to navigate complex and dynamic urban environments. Traditional motion planning approaches often struggle in environments lacking structured references or involving both static and dynamic obstacles, such as parking lots. This study proposes a semi-supervised deep transfer learning framework that enhances motion planning by enabling scenario-level knowledge transfer across diverse contexts. By leveraging successor representations, the framework captures transition relationships within specific scenarios, facilitating the adaptation of learned knowledge to new environments. This capability is incorporated through the Knowledge-based Scenario Transfer Learning and Artificial Potential Fields (KSTL-APF) approach. This study focuses on validating the proposed framework through simulation under static and semi-dynamic conditions, with real-world deployment identified as a future development phase. Experimental results indicate that KSTL-APF significantly improves planning efficiency, reduces computational time, and enhances trajectory robustness compared to conventional motion planning methods. These findings suggest that scenario-level knowledge transfer can significantly increase the adaptability and reliability of autonomous vehicle systems in unstructured or unfamiliar environments.