World Dreamer preprint reports gains in coordinated robot control

Researchers report that World Dreamer v5-lite, a model that predicts separate actions for multiple robots, achieved 78.8% mean success across four real-robot tasks, a margin of 13.8 percentage points over the strongest evaluated baseline, in an arXiv preprint. The model places calibrated camera features into a shared representation of the workspace. Each robot reads that representation and exchanges information about position and orientation while retaining its own action stream. Training with selected camera views withheld encourages robots to use observations from one another. On the RoboFactory benchmark, it achieved 74.0% mean success, 23.7 percentage points above the strongest evaluated baseline. The real-robot result covers four tasks.