Probabilistic Reachable-Action Verification of Visuomotor Policies via Set-Based Training

2026-08-03

AI Summary

This work addresses the computational expense and conservatism in reachability analysis for vision-based motion policies, which arise from high-dimensional visual encoders. The authors propose a novel framework that freezes the visual encoder and confines set propagation to a low-dimensional interface, employing zonotope-based propagation to tighten the bounding width of terminal actions. By integrating sample-based perturbations with partitioned conformal calibration, the method generates probabilistic reachable action radii with finite-sample coverage guarantees. Evaluated on closed-loop tasks, the approach significantly reduces reachable action radii while preserving task performance, outperforming baseline methods based on behavioral cloning, observation consistency, and pointwise adversarial robustness.

Abstract

Reachability analysis for visuomotor policies is difficult because large visual encoders make end-to-end set propagation computationally expensive and excessively conservative. We therefore freeze the visual encoder and confine set propagation to a low-dimensional interface between it and the downstream policy, with the interface set calibrated from held-out camera-pose perturbations. Propagating this set through the policy with zonotopes yields a terminal output-enclosure width that set-based training optimizes directly. During evaluation, camera-pose perturbations are sampled from the prescribed distribution, and rollout-level split conformal calibration converts the resulting action-deviation scores into a probabilistic reachable-action radius with finite-sample coverage. In controlled manipulation experiments, set-based training reduces this radius while preserving closed-loop task capability, and matched behavior-only, observational-consistency, and pointwise-adversarial controls all leave a larger radius.

Problem

• probabilistic reachable-action verification
• visuomotor policies
• reachability analysis
• set propagation
• visual encoders

Innovation

• probabilistic reachable-action verification
• set-based training
• zonotopes
• split conformal calibration
• visuomotor policies