Toward Human-guided Safe Reinforcement Learning in the Real World

点击翻译按钮获取中文摘要

This project aims to investigate human-guided safe reinforcement learning (RL). Safe RL is an important topic that could enable real applications of RL systems by addressing safety constraints. Existing safe RL assumes the availability of specified safety constraints in mathematical or logical forms. This project proposes to study learning safety objectives from information provided directly by humans or indirectly via language models, and human-guided continuous correction for safety improvements. The established theories and developed algorithms will advance frontier technologies in AI and contribute to a wide range of real applications of safe RL, such as robotics and autonomous driving, bringing enormous social and economic benefits.