Continual Reinforcement Learning Framework for Scalable Collision Avoidance and Mitigation System With Packing Strategy
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
Collision Avoidance and Mitigation System (CAMS) in autonomous driving systems is crucial for ensuring safety by formulating strategies to address various collisions and planning the trajectory accordingly. Although recent learning-based motion planning methods for CAMS have shown promising results for specific collision scenarios, the question of how to continually scale up their knowledge across different driving environments has not yet been thoroughly investigated. In this paper, we propose a scalable learning framework for CAMS that combines deep reinforcement learning and continual learning. By sequentially packing knowledge of new scenarios into a model trained on specific scenarios, the proposed method continually scales up the range of scenarios the model can handle while maintaining performance on previously trained scenarios. We apply this learning framework to a motion planner for CAMS and evaluate its scalability across new driving environments. This approach demonstrates a significant ability to pack knowledge by outperforming a model that jointly trains on all scenarios—typically considered an upper bound for continual learning—in about 2.4× fewer training steps, with a 4.19% higher collision avoidance rate and a 12.02% lower collision impact. Moreover, our approach achieves performance close to models trained on each scenario individually, with a model about 5.8× smaller.
Fields of Study & Subjects
Autonomous Vehicle Technology and Safety
Reinforcement Learning in Robotics
Robotic Path Planning Algorithms
Publication details
- Journal / source
- IEEE Robotics and Automation Letters
- Volume
- 11
- Issue
- 5
- Pages
- 6264–6271
- Publication date
- 27 March 2026
- Language
- EN