Title: Cross-Embodiment Imitation for Robot Whole-body Skill Learning

Date: Monday,  Nov 24th 2025

Time: 12:00 PM -  2:00 PM EST

Location: Coda C1108 Brookhaven, Zoom Link

Tianyu Li

Ph.D. Student in Computer Science

School of Interactive Computing 

Georgia Institute of Technology 

https://easypapersniper.github.io/

Committee: 

Dr. Sehoon Ha (Advisor) – School of Interactive Computing, Georgia Institute of Technology

Dr. Danfei Xu – School of Interactive Computing, Georgia Institute of Technology

Dr. Greg Turk – School of Interactive Computing, Georgia Institute of Technology

Dr. Karen Liu – Computer Science Department, Stanford University

Dr. Marco Hutter –  Robotic Systems Lab, ETH Zürich

Abstract: 
Robots need to acquire natural, human-like skills to assist humans across a wide range of tasks. A direct and intuitive approach is to learn these skills by imitating human demonstrations. However, robots often differ drastically from humans in both morphology and dynamics—ranging from quadrupeds and wheeled robots to multi-arm manipulators. This embodiment gap presents significant challenges for transferring complex, whole-body human motions to robots.

In this talk, I will present my research on enabling robots to learn sophisticated whole-body skills from human demonstrations. The talk is structured in three parts: (1) cross-morphology motion retargeting, (2) cross-embodiment interaction imitation, and (3) cross-embodiment generalization. In the first part, I will introduce methods for retargeting human motions to robots by establishing motion correspondence, addressing both cases with and without robot datasets. In the second part, I will discuss learning interaction skills from human demonstrations, including moving large and heavy objects as well as inter-agent interactions such as dancing, sparring, and handshaking. Finally, I will share my ongoing work on building foundation models for characters of arbitrary topology and on using console games as a unified platform for evaluating embodied AI alongside human performance.