ContactGuard: Pre-Contact Execution Monitoring with Action-Conditioned Latent World Models

Gehan Zheng · Matthew Johnson-Roberson · Weiming Zhi

Figure 1: Pre-contact online monitoring via action-conditioned latent world models. While a visuomotor policy executes an action chunk, ContactGuard conditions a frozen latent world model on the same planned chunk at an anchor $k$ steps bef — FIGURE FROM THE PAPER

TL;DR

ContactGuard predicts future visual latents from a policy's planned actions to abort likely contact failures before they happen, without modifying the policy.

BRIEF

Robots that manipulate objects often only discover a failed grasp after contact has already happened, when the object may have been pushed, missed, or dropped. ContactGuard checks the policy's planned action chunk just before contact. A world model trained on unlabelled robot trajectories predicts what the scene will look like shortly afterward, in a compact latent space rather than as video. A small classifier flags likely failure and aborts. On real contact-rich tasks it predicted failure more accurately than ablations, without changing the policy.

Abstract

Contact-rich manipulation failures are often detected only after the robot has committed to contact. This is especially limiting in wrist-camera setups: close gripper–object views help observe contact, but a poor approach may already push, miss, slip, or disturb the object before conventional detectors react. We introduce ContactGuard, a pre-contact execution monitor for chunked visuomotor policies. Given the policy's planned action chunk, ContactGuard predicts its short-horizon consequence in latent visual space and aborts if the predicted future latent indicates likely failure. Its latent world model is trained from unlabelled robot trajectories to predict compact multi-view visual embeddings under planned actions, avoiding pixel-level video prediction. A lightweight failure probe is then trained from a small labelled set of pre-contact clips. At deployment, ContactGuard anchors prediction before an imminent contact event, rolls the model forward under the policy's own actions, and verifies the predicted post-contact latent. Across real-world contact-rich manipulation tasks, ContactGuard predicts failure more accurately than direct and corrupted-action ablations, and transfers to live robot as a pre-contact abort signal without modifying the underlying policy.