A mosaic of humanoid robots interacting with boxes, chairs, luggage, tools, and other objects.

A data flywheel for humanoid loco-manipulation.

InterMimicGen

Scaling humanoid loco-manipulation through self-evolving motion imitation.

Yucheng Zhang1† Sirui Xu1†* Jinhong Li1† Liuyu Bian1† Anatulya Nandi1 Derek Zhang1 Xiangchen Liu1 Xueting Li2 Umar Iqbal2 Yu-Xiong Wang1‡ Liang-Yan Gui1‡

1University of Illinois Urbana-Champaign2NVIDIA†Equal Contribution*Project Lead‡Equal Advising

University of Illinois Urbana-ChampaignNVIDIA
Paper on arXiv Explore the research
Scroll to explore

01 / Deployment

InterMimicGen
for physical robots.

The idea

One human interaction can become many robot motions.

Scale like MimicGen, ground like InterMimic: every physically executed motion becomes new data.

02 / Retargeting

InterMimicGen
for resolving the embodiment gap.

02.1

Across objects

02.2

Across embodiments

02.3

Against baselines

03 / Tracking

InterMimicGen
for large-scale motion imitation.

03.1

Across objects

03.2

Across robot configurations

04 / Evolution

InterMimicGen
for self-evolution.

Local edits.
Growing reach.

Accepted physical executions reach farther from the original interaction over successive rounds.

Accepted object placements expand outward from the original seed over five augmentation rounds.

One motion, many realizations.