ContactGuard: Pre-Contact Execution Monitoring with Action-Conditioned Latent World Models
ContactGuard is a pre-contact execution monitor for chunked visuomotor policies. It predicts short-horizon consequences of planned action chunks 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 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.
ContactGuard introduces a pre-contact execution monitor for chunked visuomotor policies, designed to detect likely failures before a robot commits to contact. It uses an action-conditioned latent world model trained on unlabelled robot trajectories to predict multi-view visual embeddings under planned actions, and a lightweight failure probe trained on a small labelled set of pre-contact clips. In real-world contact-rich manipulation tasks, ContactGuard predicts failure more accurately than direct and corrupted-action ablations.
The method avoids pixel-level video prediction by operating in a compact latent visual space, using multi-view embeddings. It leverages the policy's own planned action chunk to roll forward the world model, enabling pre-contact verification. The failure probe is trained on a small labelled set, suggesting efficient use of supervision. The reported improvement over direct and corrupted-action ablations indicates the value of action-conditioned prediction for early failure detection.
Pre-contact failure detection can reduce damage and improve reliability in robotic manipulation, especially in wrist-camera setups where close views are limited. This approach may lower the cost of deploying contact-rich manipulation by reducing trial-and-error and enabling safer abort mechanisms. The use of unlabelled trajectories for world model training suggests scalability to diverse tasks without extensive labelling.
Improved failure prediction can reduce downtime, damage, and wasted cycles in industrial manipulation, potentially lowering operational costs and increasing throughput. The method's reliance on unlabelled data and small labelled sets may reduce data collection costs for deployment.
Potential next signals include publication of code or datasets, follow-up work on extending to other robot embodiments or sensor modalities, and integration into commercial robot manipulation stacks. Evaluation on additional contact-rich tasks or comparison with other pre-contact monitoring methods would further validate the approach.