[Invited Talk] Control and Path Planning of Walking Robots Using Reinforcement Learning

Presenter: Jemin Hwangbo (KAIST); Time: 4:00pm, Thursday (2022/06/23); Location: Room 204, Building 133

Abstract

In this talk, research on the design and control of a walking robot conducted by KAIST RaiLab will be introduced. RaiLab has recently developed a quadruped robot, Raibo, and is using it for various control studies. In particular, research is in progress on network structure, dynamic modeling, and learning framework construction in order to safely walk in a complex environment using reinforcement learning. It is characterized by controlling with the sim-to-real technique used in real robots by learning in RaiSim, a dynamic simulator developed by ourselves. These studies enable robust control of a quadruped robot that jumps on sand and does not fall over in spite of large disturbances. Also, in this presentation, an example of easy and fast reinforcement learning using Raisim GymTorch, an Open Source Reinforcement Learning Framework, will be explored.

Biography

Professor Jemin Hwangbo is currently an assistant professor in the Department of Mechanical Engineering at KAIST. He obtained his Ph.D. from the Department of Mechanical Engineering at Zurich University of Technology in 2019, and until 2020 conducted a postdoctoral research program at the Robot Systems Laboratory (RSL) at Zurich University of Technology under the guidance of Professor Marco Hutter. He is actively conducting research activities in the field of learning-based multi-joint robot control using deep models and developing a physics engine for simulation. His representative papers include “Learning agile and dynamic motor skills for legged robots” published in Science Robotics, "Control of a quadrotor with reinforcement learning," and "Per-contact iteration method for solving contact dynamics" published in IEEE RA-L. He is also distributing and developing a cross-platform physics engine simulator called RaiSim, which improves computational accuracy and performance with a groundbreaking contact solver method compared to conventional simulations.