Interactive Reinforcement Learning from Natural Language Feedback

摘要

Large Language Models (LLMs) are increasingly influential in advancing robotics. This paper introduces ECLAIR (Evaluative Corrective Guidance Language as Reinforcement), a novel framework that leverages LLMs to interpret and incorporate diverse natural language feedback into robotic learning. ECLAIR unifies various forms of human advice into actionable insights within a Reinforcement Learning context, enabling more efficient robot instruction. Experiments with real-world users demonstrate that ECLAIR accelerates the robot’s learning process, aligning its policy closer to optimal from the outset and reducing the need for extensive human intervention. Additionally, ECLAIR effectively integrates multiple types of advice and adapts well to prompt modifications. It also supports multilingual instruction, broadening its applicability and fostering more inclusive human-robot interactions. Project website: https://sites.google.com/view/eclairiros