Shi Fan

I am a Postdoctoral Researcher at ETH Zürich working with Prof. Stelian Coros and Prof. Marco Hutter.

I obtained the PhD degree at JSK Lab, the University of Tokyo, supervised by Prof. Masayuki Inaba and Prof. Kei Okada from 2016 to 2021. In 2020, I was visiting RSL Lab, ETH Zurich supervised by Prof. Macro Hutter.

I did my Bachelor in Peking University advised by Prof. Huijing Zhao. I was visiting in Microsoft Research Asia (by Prof. Katsushi Ikeuchi) and Takanishi Lab (by Prof. Atsuo Takanishi) during my undergraduate.

In spare time, I was fortunate to work with Erico Guizzo and Evan Ackerman as a Contributor on IEEE Spectrum Robotics, and my great friend Yifan Hou on 机器人学家. My lifelong target is to build the good stuff, and help the good happen.

Website: https://fanshi14.github.io/me/

Reference-based vs Reference-free Reinforcement Learning for Humanoid Robot

Abstract

Reinforcement learning recently shows great progress on legged robots, while humanoid robots in high dimensions but narrow solution space are still challenging to learn. The typical methods introduce the reference joints motion to guide the learning process, which we called reference-based method; however, obtaining a high-quality reference trajectory is nontrivial, and imitation suffers from the local minimum. For general reference-free scenarios, the humanoid robot is discouraged by the early termination and biased sample collection. In this talk, we will introduce our recent progress in both reference-based and reference-free methods for humanoid robot. We try to compare them and provide more discussion in learning-based approach on humanoid robot control.