ContactFlow: A video action conditioning that transfers across embodiments
For robot learning, this work addresses the bottleneck of embodiment-specific action conditioning by proposing a representation that generalizes across humans and robots, enabling scalable world model training from diverse data sources.
ContactFlow introduces an embodiment-agnostic action representation based on 3D contact point trajectories, enabling a video generative world model trained on both human and robot data to predict physically plausible manipulation outcomes. In experiments on DROID and real-world tasks, it enables transfer between human demos and different robotic embodiments.
World models offer a promising route toward robot planning by enabling agents to imagine and verify the consequences of actions before execution. However, current video-based world models often struggle to capture the physical constraints that govern manipulation, particularly contact. Further, their action conditioning is often constrained to specific embodiments such as parallel grippers. We propose \emph{Contact Flow}, an embodiment-agnostic action representation that encodes manipulation through the trajectory of 3D contact points between an actor and a target object. By discarding actor-specific appearance and kinematics, Contact Flow provides a shared conditioning signal for both human demonstrations and robotic execution. Therefore, we can train a large-scale video generative model on both human and robotic object interaction videos conditioned on Contact Flow, yielding a world model that predicts physically plausible manipulation outcomes. We integrate this model into a propose-imagine-verify-act pipeline, where generated rollouts are assessed by a vision-language model before execution. Experiments on the DROID dataset and real-world tabletop manipulation tasks demonstrate that Contact Flow enables transfer between human demonstrations and different robotic embodiments.