Researchers at Nanyang Technological University and Chongqing University developed a continual learning method that trains self-driving vehicles using intervention data from human drivers, according to a study published in Nature Communications. The technique uses data gathered when human drivers take over the wheel during ambiguous traffic situations and rare long-tail events.
The system updates driving policies using only the new data collected while driving, avoiding the need for lengthy retraining from scratch. In both simulation tests and real-world experiments, the updated policy matched or outperformed earlier driving software on social compliance and handling rare cases.
Autonomous vehicles frequently encounter edge cases where road laws are unclear or unusual conditions occur, forcing human operators to intervene. Rather than discarding those takeover events, the team’s approach incrementally incorporates small amounts of human guidance to prepare the vehicle for similar encounters in future trips. The researchers noted that the approach could support broader human-in-the-loop autonomous systems across multiple operational dimensions.
Haohan Yang and Yanxin Zhou contributed equally as lead authors, working alongside Haochen Liu and Chen Lv at Nanyang Technological University in Singapore, as well as Xiaosong Hu and Chaoyue Chen at Chongqing University in China. The study was received on July 30, 2025, accepted on August 10, 2026, and published on September 5, 2026. Funding support came from Singapore’s Agency for Science, Technology and Research and Ministry of Education, as well as the National Natural Science Foundation of China and the Basic Research Funds for Central Universities.
