Event

World Dreamer v5-lite Multi-Robot World Action Model Released

1 reports · 1 sources

Guoning Wu and other authors released World Dreamer v5-lite, a world action model for multi-robot settings: it uses a shared world state to predict action sequences separately for each robot, with each robot's action stream remaining independent. The model builds the world state in workspace coordinates from calibrated features of a pretrained video encoder, and its spatial size does not change as the number of viewpoints increases; a vision-language model provides scene context and supplies one query per robot, with each query reading the shared state and exchanging pose-aware information. Pretraining used over 10,000 hours of trajectories from 20 types of simulated robots, RoboFactory, and real robots; training employs masked-view coordination, withholding selected camera views from the two visual branches while retaining all action targets. The authors report an average success rate of 74.0% on RoboFactory, 23.7 percentage points higher than the strongest evaluated baseline, and an average success rate of 78.8% on four real-robot tasks, leading by 13.8 percentage points.