FoMo-FD catches 96.6% of surgical robot failures

Surgical robots trained via imitation learning face a critical safety hurdle: knowing when they're about to screw up. Existing failure detection systems struggle with scarce failure data, unpredictable manipulation dynamics, and the tradeoff between catching mistakes and avoiding disruptive alarms. A new paper introduces FoMo-FD (Flow-Matching World Model for Failure Detection), which sidesteps the need for failure examples entirely. FoMo-FD learns what normal short-horizon visual dynamics look like using an action-conditioned flow-matching world model, then scores the inverse-transport nonconformity between observed endpoint latents to detect when the robot's actions diverge from expected outcomes. Detection thresholds are set via conformal calibration on successful executions, giving task-specific alarms without assuming future failure types.

The method was evaluated on four surgically relevant manipulation tasks with twenty failure modes across simulation and real-world experiments using the da Vinci Research Kit (dVRK). Results show FoMo-FD outperforms observation-level anomaly baselines and a prediction-error variant of the same world model. The wrist-camera view delivered the strongest performance: 96.6% failure detection rate at a 1.3% false alarm rate. This means surgeons and supervisors could trust the system to catch nearly all potential errors while only rarely interrupting a well-executed procedure. By learning dynamics from successful runs alone, FoMo-FD offers a practical safeguard for autonomous surgical systems, where collecting failure demonstrations is often impractical or unsafe. The paper, submitted to IEEE Robotics and Automation Letters, marks a step toward safer deployment of imitation learning in surgery.