Bimanual manipulation is hard for world models because the two arms act on the same scene while each wrist camera sees only its own view. Existing interactive world models typically condition on text or a single camera, so they struggle to follow per-arm actions and keep the two views consistent. UMI-WM is designed to address this by jointly generating both wrist views from observation history, per-arm relative camera poses, and gripper widths.

The model routes each arm's action into its own wrist stream, and sparse hub tokens pass scene information between the two streams. Geometry conditioning tied to camera parameters and camera-guided noise help control motion per view and support geometric exchange across views. The authors also adapt the model for block-causal generation and distill a 30-step teacher into a student that uses only four denoising steps.

Training data combines self-collected open-world and in-studio UMI recordings with public manipulation data. To evaluate action following, the authors use an inverse dynamics model to recover camera poses and gripper widths from generated videos, and introduce a Bimanual Action Fidelity (BAF) metric. UMI-WM reports the best results on five reference-video fidelity metrics and competitive performance on VBench and WorldArena; the four-step student generates paired wrist videos 6.9 times faster than the teacher.