Researchers at the University of Illinois Urbana-Champaign and NVIDIA report that InterMimicGen, a system for turning captured human-object interactions into robot motion references, expanded its verified motion library to more than 140 times its original size after five rounds in simulation. That growth reflects new variations of tasks already represented in the captured interactions; it does not show that a robot learned to perform an open-ended range of new jobs.
The team combined 16,059 motion-capture sequences from the InterAct and HiPHI datasets, then retargeted the body movements and hand-object contacts to humanoid robots. A physics-based tracker tries to follow each reference. In each round, the system makes small changes to object placement or body posture, fine-tunes the tracker and keeps simulated runs that complete the task without a sustained fall and pass contact and motion-quality checks. Accepted variations seed the next round.
After five rounds, the verified reference set grew to 150.5 times its starting size for bimanual tasks on a Unitree G1 with Inspire hands, 142 times for bimanual tasks on a G1 with Dex3 hands, and 146.4 times for grasping tasks on the Inspire setup. In a separate replay test, a reference counted as executable if at least one of three attempts reached its final frame. The original trackers completed 52% to 64% of generated references; the evolved trackers completed 98% to 99%. Both versions still completed all of the original references in that test.
The expansion came with a quality trade-off: foot-sliding frames rose from roughly 1% to roughly 2% of contact frames across the three settings. The paper also shows selected motions transferred to physical robots, but its five-round scale results are simulated and its variations remain anchored to captured interactions. The study is an arXiv preprint.
Sources
- InterMimicGen: Scaling Humanoid Loco-Manipulation through Self-Evolving Motion Imitation (arXiv:2610.06850v1, October 5, 2026)
- Illustrative photo: RuinDig / Yuki Uchida, Unitree G1 at Japan Mobility Show 2025, licensed under CC BY 4.0. The exhibition photo is not a study trial.
