Learning motor skills for robot locomotion and manipulation - Zhibin (Alex) LI, Director of AIR Lab, University of Edinburgh
Date: Wednesday 1 September 2021 14:30 - 15:15
Location: MS TEAMS
Multi-expert learning of adaptive legged locomotion: https://youtu.be/n34gGhb7Qjo
Abstract: One of the overarching scientific goals of robot control of physical interactions is to develop machines to traverse various terrains, deliver payload, and perform manipulation & grasping tasks with a mixed level of autonomy: from remote control to fully autonomous operation. This talk will cover primarily on the "deep reinforcement learning" for continuous robot control in recent years and discuss how innovation of these domains that can make a step change for solving real-world problems. Particularly, I will cover specific techniques on multi-skill learning of locomotion and manipulation and a brief discussion on sim2real.
I will showcase some new results from deep reinforcement learning combined with control techniques in a hierarchical framework to self-learn goal-oriented policies on various locomotion, dexterous manipulation & grasping tasks. I will elaborate on the multi-expert learning for fusing expert skills which showcases the new capabilities to achieve fall recovery, trotting, target-following and any transitional skills. I will also present the meta-learning approach, how the SpotMicro robot can produce adaptive and robust locomotion skills under changing ground friction, external pushes, and different robot dynamics including motor failures and the whole leg amputation – all done within 0.2 s (like in the robot dog in “Black Mirror”).
In the end of the talk, we will embrace in-depth and interactive discussions about future research directions, and how researchers can collaborate and explore new research areas.
Dr Zhibin (Alex) Li is an Assistant Professor at the University of Edinburgh, Head of the Advanced Robotics Intelligence Laboratory, and a member of the Edinburgh Centre of Robotics. He has published more than 80 papers in robotics, such as Science Robotics and IEEE Transaction on Robotics. His recent work of "multi-expert learning of adaptive legged locomotion" is the first implementation of adaptive quadrupedal locomotion on a real robot using a multi-expert learning architecture (selected as a cover article in the December 2020 issue of Science Robotics).
He is an associate editor of IEEE RAL Letters, and his research interest covers dynamic motion skills of legged robots, optimization-based robot motion planning and control, and deep reinforcement learning for robot motor learning including locomotion, manipulation and grasping. He is leading the theme of "Shared and Autonomous Manipulation" in joint cross-hub demonstration of the UK National Robotics and Artificial Intelligence Hubs, and is working on developing novel optimization and learning based robot control to achieve a new level of robot autonomy and intelligence.