Human hands
Learn interaction structure from abundant human demonstrations.
ContactFlow is an embodiment-agnostic action representation that describes manipulation through trajectories of 3D contact points— shared by hands, grippers, and robots.
The same physical intent can look radically different across embodiments.
Contact is the common language.01 / The idea
Instead of conditioning a world model on actor-specific appearance or joint commands, ContactFlow isolates where interaction happens and how that interface moves through 3D space.
Learn interaction structure from abundant human demonstrations.
Express robot plans using the same contact-centric signal.
Transfer action conditioning without copying actor geometry.
02 / Preview
The model receives an initial scene and a ContactFlow condition, then predicts the visual outcome. Each six-second result explicitly pairs our generated future with its matching ground truth.
Close-up bimanual interaction with clear object motion.
Egocentric interaction with a clearly visible manipulated object.
Cross-embodiment robot interaction in a cluttered real scene.
High-quality held-out human interaction in a real kitchen scene.
Clear held-out tool interaction selected for the paper.
Genuine 49-frame paper results. Every pair is synchronized and explicitly labeled—our generated rollout first, matching ground truth second.
03 / Release
This is the official release repository. We will update it with the paper, implementation, checkpoints, and processing tools as they become available.
@inproceedings{contactflow2026,
title = {ContactFlow: A Video Action Conditioning
that Transfers Across Embodiments},
year = {2026}
}
Complete author and venue metadata will be added after review.