Scenario-level knowledge transfer for motion planning of autonomous driving via successor representation

For autonomous vehicles, transfer learning can enhance performance by making better use of previously learned knowledge in newly encountered scenarios, which holds great promise for improving the performance of motion planning. However, previous practices using transfer learning are data-level, which is mainly achieved by introducing extra data and expanding experience. Such data-level consideration depends heavily on the quality and quantity of data, failing to take into account the scenario-level features behind similar scenarios. In this paper, we provide a scenario-level knowledge transfer framework for motion planning of autonomous driving, named SceTL. By capitalizing on successor representation, a general scenario-level knowledge among similar scenarios can be captured and thereby recycled in different traffic scenarios to empower motion planning. To verify the efficacy of our framework, a method that combines SceTL and classic artificial potential field (APF), named SceTL-APF, is proposed to conduct global planning for navigation in static scenarios. Meanwhile, a local planning method combining SceTL and motion primitives (MP), SceTL-MP, is developed for dynamic scenarios. Both simulated and realistic data are used for verification. Experimental results demonstrate that SceTL can facilitate the scenario-level knowledge transfer for both SceTL-APF and SceTL-MP, characterized by better adaptivity and faster computation speed compared with existing motion planning methods.