Learning from Mistakes: Rollout-Retrieval Lifelong Policy Learning for Autonomous Driving
从错误中学习:面向自动驾驶的展开-检索终身策略学习
机构 * School of Mechanical Engineering, Beijing Institute of Technology(北京理工大学机械工程学院) ; School of Mechanical and Aerospace Engineering, Nanyang Technological University(南洋理工大学机械与航空航天工程学院)
AI总结 提出R²LPL框架,通过从可恢复的闭环驾驶错误中检索纠正目标并进行终身学习,将稀疏失败证据转化为监督知识,持续提升自动驾驶策略性能。
Comments 15 pages, 6 figures. Code available at: https://github.com/Engibacter/R2LPL