Researchers unveiled World Dreamer v5‑lite, a model that predicts separate action sequences for each robot while drawing on camera views from many robots. The system creates a shared world state that holds calibrated video features in workspace coordinates. All robot queries read this common state, exchange pose‑aware information, yet keep distinct action streams. To make robots rely on shared observations, the training process masks random camera feeds during learning but still gives all robots their real action targets.

On the RoboFactory benchmark, World Dreamer v5‑lite achieved a 74.0 % mean success rate, beating the strongest baseline by 23.7 percentage points. For four real‑robot tasks, the model reached a 78.8 % mean success, 13.8 percentage points higher than the best baseline. On the RoboCasa365 benchmark, its overall success rate was 58.08 %, surpassing the top leaderboard method (PatchWAM) by 10.48 percentage points.