PRISM turns 4 human videos into 256 variants to train generalist humanoid loco-manipulation policy
GordonWetzstein · x · 2026-11-02
Researchers at Amazon FAR (with Pieter Abbeel, Rocky Duan, Jitendra Malik and others) will present PRISM at CoRL 2026, bringing rich human interaction experience to humanoid loco-manipulation via scalable real-to-sim-to-real.
- Only 4 real human videos are expanded into 256 counterfactual variants using a video generation model
- Robot-object trajectories are recovered from the videos
- A single policy trained on this data generalizes across diverse objects, scales, and spatial layouts
Jitendra Malik hailed it as another example of the power of human videos as a data source, declaring that the real2sim2real paradigm rules.
More from Embodied
- Dev recreates 'subvocal' for AirPods Pro 3 with one prompt in 9 minutes, open-sourced — _AustinCalvert_ · 2026-11-02
- Scaling past 100 robots without an army of field engineers: 5 ops practices — StewartalsopIII · 2026-11-02
- Neuralink Patient 22 Demonstrates Controlling Robots Telepathically in New Video — Scobleizer · 2026-11-02
- Little grok bot takes its first wobbly steps, then gives up abruptly — Baconbrix · 2026-11-02
- Two UIST 2026 papers in one demo: thermal cold illusion and haptic touch ring — plopesresearch · 2026-11-02
- Wheel-free cleaning robot hangs from 4 points, directed by Claude Opus 5.5 — LinusEkenstam · 2026-11-02