Counterfactual Video Generation Enables Scalable Humanoid Loco-Manipulation

arXiv:2609.38172 · cs.RO, cs.CV, cs.GR · Submitted 2026-09-29 · Read on arXiv

cs.RO, cs.CV, cs.GR

Submitted: 2026-09-29

Updated: 2026-09-29

Comments: published at CoRL 2026. Project page: https://prism-real2sim2real.github.io/

License: http://creativecommons.org/licenses/by/4.0/

The gist: Teaching humanoids loco-manipulation skills, such as carrying diverse objects, via visual imitation is a promising path toward generalist robots.

Terminology

Abstract

Teaching humanoids loco-manipulation skills, such as carrying diverse objects, via visual imitation is a promising path toward generalist robots. However, collecting diverse, high-quality interaction videos, such as clips that clearly show a person's full body and unoccluded interactions with objects, poses a practical barrier to scaling this approach. We propose PRISM, a real-to-sim-to-real framework that overcomes this limitation by amplifying a handful of real videos into a large, diverse training set. PRISM first generates hundreds of diverse "counterfactual" human-object interaction videos via video-to-video (V2V) generation from a few exemplar real videos. Our contact-anchored real-to-sim pipeline then reconstructs both human and object motions, retargeting this imperfect video data into physically plausible trajectories. The intra-class variability across these counterfactual videos lets us train a single policy that generalizes to unseen objects within each category. We demonstrate the full pipeline by deploying this policy on a real robot without any real-world fine-tuning. Using only onboard depth observations, our humanoid picks up, carries, and drops objects, including boxes, barrels, bins, and balls, across novel instances, sizes, and initial configurations.

Sources

Related papers