Fall 2022 GRASP SFI - Zhongyu Li, University of California Berkeley скачать в хорошем качестве

Fall 2022 GRASP SFI - Zhongyu Li, University of California Berkeley

ABSTRACT In this talk, I will provide a brief introduction about our recent progress in applying optimal control and deep reinforcement learning (RL) on legged robots in the real world. I will then dive into our recent work to bridge model-based safety-critical control and model-free RL on a highly nonlinear and complex system, such as a bipedal robot Cassie. Bridging model-based safety and model-free RL for dynamic robots is appealing since model-based methods are able to provide formal safety guarantees, while RL-based methods are able to exploit the robot agility by learning from the full-order system dynamics. I will discuss a new method to combine them by explicitly finding a low-dimensional model of the system controlled by a RL policy and applying stability and safety guarantees on that simple model. PRESENTER Zhongyu Li is a fourth-year PhD student in Mechanical Engineering at UC Berkeley. He is advised by Prof. Koushil Sreenath and focuses on optimal control and reinforcement learning (RL) for legged robots. His work has enabled a bipedal robot Cassie to perform robust and agile maneuvers and to navigate autonomously in unknown and cluttered environments. His work has also enabled quadrupedal robots to function as guide dogs, soccer ball shooters, and collaborative agents. Zhongyu’s work has been the Best Entertainment and Amusement Paper Finalist (IROS 2020), the Best Service Robot Paper Finalist (ICRA 2021), and the Best RoboCup Paper Finalist (IROS 2022).