Action-Free Video Pretraining Lifts Surgical Robot Success From 63.5% to 77.8% Under a Fixed Demonstration Budget

arXiv 2608.11204·medium signalXBlueskyLinkedInCopy link

Surgical WAM tests whether cheap action-free endoscopic video can substitute for scarce action-labeled teleoperated trajectories, which require synchronized kinematics and are costly to collect. Built on Cosmos Policy, it jointly predicts future endoscopic observations and executable action chunks — first learning visual dynamics from action-free video, then fine-tuning on a fixed action-labeled budget, and at deployment acting as a receding-horizon controller that executes a short prefix of each chunk and replans. Across four simulated tasks, video pretraining raised average success from 63.5% to 77.8%, with a 20-point absolute gain on PegTransfer and the largest improvements on contact-rich and bimanual tasks.