Reinforcement Learning in Robotics: Wins, Fails, and the Next
Reinforcement learning (RL) has demonstrated remarkable successes in robotic domains ranging from agile drone racing to complex locomotion and dexterous manipulation. Yet, its widespread deployment in real-world robotics remains constrained by key challenges, including sample inefficiency, the sim-to-real mismatch, and the untransparent nature of black-box models. In this talk, I’ll also share our recent work on simulation, differentiable learning methods, and AI safety, aiming toward the development of more intelligent, robust, and trustworthy robotic systems.
BIO: Fan Shi is an Assistant Professor at the National University of Singapore (NUS) awarded by NUS Presidential Young Professorship, where he leads the Human-Centered Robotics Lab. He did his Postdoc in ETH Zurich and PhD in the University of Tokyo on learning-based legged locomotion. Now his research focuses on developing intelligent and safe physical AI agents to advance robot deployment in human-centered environments.