ContactGuard: Pre-Contact Execution Monitoring with Action-Conditioned Latent World Models
The monitor tries to stop a robot before a bad contact happens, using the policy’s own planned actions to predict the near future.
ContactGuard predicts short-horizon outcomes in latent visual space, then aborts if the projected post-contact state looks like a likely failure. The model is trained on unlabeled robot trajectories, with a smaller labeled set used for the failure probe. The authors report better failure prediction than direct and corrupted-action ablations across real-world contact-rich manipulation tasks, and say it transfers to a live robot without changing the underlying policy. ArXiv · AI/CL/LG's note
ContactGuard predicts short-horizon outcomes in latent visual space, then aborts if the projected post-contact state looks like a likely failure. The model is trained on unlabeled robot trajectories, with a smaller labeled set used for the failure probe. The authors report better failure prediction than direct and corrupted-action ablations across real-world contact-rich manipulation tasks, and say it transfers to a live robot without changing the underlying policy. ArXiv · AI/CL/LG's note
score 5