InterPReT：交互式策略重构与训练赋能非专业人员高效模仿学习

Imitation learning holds great promise for robotics, yet learning from layperson demonstrations presents significant challenges due to suboptimal demonstration quality, poorly structured policies, and the absence of effective feedback mechanisms. This paper introduces InterPReT, an interactive policy restructuring and training framework designed to address these issues. InterPReT empowers users to review and modify the policy structure post-demonstration, enabling them to add or remove subtasks and provide corrective feedback for erroneous behaviors. The framework leverages large language models (LLMs) to translate user feedback into actionable policy modifications and employs reward-based reinforcement learning (RL) for policy optimization. Experimental results demonstrate that InterPReT substantially improves the success rate and efficiency of learning from layperson demonstrations across a variety of robotic manipulation tasks. A user study further validates its usability and effectiveness, empowering non-expert users to easily teach complex skills to robots. This advancement facilitates broader real-world applications of imitation learning by making it more accessible and robust, ultimately bridging the gap between human intent and robotic execution in practical scenarios.