PsiBot's key trick is to reverse the usual direction: instead of translating human footage into robot data, the Psi-W0 world model inversely generates human-hand data from existing robot trajectories. A video-editing-style transformer — not a policy — is then trained to map human-hand input to aligned robot output. This lets abundant robot data synthesize matched human demonstrations rather than the other way round (the inverse of the classic sim2real approach). Quality is validated on two tracks: direct replay of the actions on a physical robot, and downstream fine-tuning success.
Human footage is cheap and plentiful, but cannot be used directly to train a robot: appearance (human hand vs gripper) and motion kinematics differ. Loosely collected data with matching intent is not enough, because the model has to infer the mapping itself. Strong pair data removes that guesswork by supplying an explicit, synchronized human→robot mapping.
The human demonstration and the corresponding robot trajectory are synchronized at the level of individual frames, not merely the whole task.
The scene is the same in both recordings, excluding differences that stem solely from the differing embodiment.
The recorded actions must replay directly and successfully on the robot — this is simultaneously the definition and the quality test for the pair.
The Psi-W0 world model generates human-hand data from robot trajectories, after which a video-editing-style transformer closes the human→robot mapping.
Manually collecting synchronized human–robot pairs is costly and scales poorly; without world-model synthesis the requirement easily becomes a bottleneck.
A pair that satisfies temporal alignment and scene consistency but whose actions the robot cannot replay correctly is not a strong pair — skipping the replay test poisons the dataset.
PsiBot described strong pair data in the technical blog accompanying the Psi-R2.5 launch, presenting it as the core of the rebuilt training pipeline.
