Ropedia just released W0, an open world model trained on ~10,000 hours of egocentric human video.

One set of weights predicts first-person views for humans, robot arms, and humanoids (Unitree G1, Fourier GR-1) using a shared camera + skeleton representation. Adding a new robot only requires its URDF.

It leads Cosmos 3 and DreamDojo across every benchmark tested. After distillation it runs at 21.3 FPS, enough for real-time VR control, and can roll out minute-long continuous loco-manipulation.

Human pretraining transfers: the model with zero robot data already beats robot-trained baselines on several humanoid tests. In practice it can rank policies, pick better actions at runtime, and transfer to new robots without large-scale data collection.

Paper, code, weights, and demos are all out.