Ropedia releases open W0 world model trained on 10,000 hours of egocentric human video
liuziwei7 · x · 2026-10-21
Ropedia has open-sourced W0, a world model trained on 10,000 hours of egocentric human video. Using a shared camera + skeleton representation, a single set of weights predicts first-person views for humans, robot arms, and humanoids (Unitree G1, Fourier GR-1); adding a new robot only requires its URDF.
Key points:
- Outperforms Cosmos 3 and DreamDojo across every benchmark tested;
- Runs at 21.3 FPS after distillation, enough for real-time VR control and minute-long continuous loco-manipulation;
- Human pretraining transfers: with zero robot data it already beats robot-trained baselines on several humanoid tests;
- Useful in practice for ranking policies, picking better actions at runtime, and transferring to new robots without large-scale data collection.
Paper, code, weights, and demos are all available.
Related event: Ropedia Open-Sources W0 World Model Trained on 10K Hours of Human Video(11 posts)→
More from Embodied
- ZJU's VLA-Rolling reuses discarded action tails to cut VLA inference cost across 16 model-benchmark pairs — OmniAI-ZJU · 2026-10-21
- Maker 3D-prints an adapter to mount a chuck onto a robot motor — _Stocko_ · 2026-10-21
- Phonon TTS pipeline ported to GenieX, targeting on-device NPU inference — HowDevelop · 2026-10-21
- Characterizing a cheap SG90 servo with Claude's plan to build a MuJoCo model — _Stocko_ · 2026-10-21
- Berkeley researchers introduce Advantage Guidance, an RL recipe for VLAs hitting ~70% success — berkeley_ai · 2026-10-21
- The New Yorker: We're Learning to Live With Waymos, and It's Going Weird — njrc · 2026-10-21