What's this about?
The world model predicts videos conditioned on robot actions, calibrates trajectories and learns from modified action scenarios. Authors report human-rated interaction defects falling from 48.12% to 6.25% on AgiBot; this evaluates generated predictions, not deployed robot safety.
What exactly is counterfactual?
Asking what would happen if a different action were taken—like replaying a missed catch with a different hand position.
What's the community saying / why does it matter?
HF: 14 upvotes; no substantive discussion collected. Seeing unsuccessful interactions can prevent a simulator from unrealistically predicting success every time.