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

2026-08-13

AI Summary

This work addresses the challenge of timely failure detection in contact-rich robotic manipulation, where existing methods typically identify failures only after contact occurs—particularly problematic under wrist-mounted camera views. To enable proactive intervention, the authors propose ContactGuard, a pre-contact execution monitor tailored for modular visuomotor policies. ContactGuard leverages an action-conditional implicit world model to predict short-horizon future visual embeddings in latent space and employs a lightweight failure detector to assess risk, allowing high-risk executions to be aborted in advance without modifying the underlying policy. By avoiding pixel-level video prediction and requiring only minimal annotated data, the approach significantly outperforms direct prediction and perturbation-based baselines in real-world tasks, achieving effective pre-contact failure warning and real-time deployment.

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 \emph{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.

Problem

• contact-rich manipulation
• failure detection
• pre-contact monitoring
• visuomotor policies
• robotic manipulation

Innovation

• pre-contact monitoring
• latent world model
• action-conditioned prediction
• visuomotor policy
• failure detection