NVIDIA's ThinkWorld learns robot skills from unlabeled human videos via chain-of-dynamics reasoning
nvidia · hf · 2026-10-28
NVIDIA and KAIST published ThinkWorld, a method for improving robotic generalization.
Key idea: represent actions as dynamics chains, align human egocentric video and robot demonstrations in a unified motion space, and use "world imagination" to reason chain-of-dynamics about future states — enabling skill learning from unlabeled human videos (EgoDex, EgoVerse, DROID) without per-clip annotation.
Experiments show it beats Diffusion Policy, MolmoAct2 and LaVAM both in-domain and out-of-domain (new objects/trajectories/environments), reaching 88.9 vs baselines' 44-46 out-of-domain, and supports skill composition from human videos.
Related event: NVIDIA and KAIST Unveil ThinkWorld: Robots Learn Skills from Human Videos(2 posts)→
More from Embodied
- Houston Robotaxi called most 'fake' AV service map: 40 min to cross 2 blocks — reed · 2026-10-28
- Indie roboticist gets first anthropometry-based robot walking with agent-driven design loop — IanPritchard · 2026-10-28
- Reward AI's OM-1 Learns Manipulation Skills From Human Demonstrations for Cross-Body Transfer — CyberRobooo · 2026-10-28
- 4 Stanford freshmen build assistive walker robot in 36 hours, win grand prize at CalHacks 13.0 — Scobleizer · 2026-10-28
- Just Action Model: decoupling what separates VLA from world-action models — Yonghao Yu · 2026-10-28
- NVIDIA Research releases Foundation4D: unified 4D perception model turns camera video into 3D scenes, no lidar needed — Scobleizer · 2026-10-28