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

Noosaga Placements

• Robot Learningsubfield90%
  The paper focuses on training a policy (set-based training) to improve verifiability, which is a robot learning contribution.
  Set-based training makes the size of a propagated output enclosure an explicit training objectiveset-based training reduces this radius while preserving closed-loop task capability
• Safe Robot Learningframework80%
  The paper's set-based training aims to improve safety and verifiability of robot policies, aligning with safe robot learning.
  Set-based training makes the size of a propagated output enclosure an explicit training objectiveprobabilistic reachable-action radius
• Robot Controlsubfield70%
  The work addresses action deviations and safety in control, with a focus on reachable-action verification.
  the largest resulting action deviation is the safety concernprobabilistic reachable-action radius
• Probabilistic Robot Learningframework70%
  The method uses probabilistic guarantees via conformal calibration and set propagation, fitting probabilistic robot learning.
  rollout-level split conformal calibration converts the resulting action-deviation scores into a probabilistic reachable-action radiuszonotopes
• Ai Safetysubfield60%
  The paper provides verification and robustness guarantees for policies under perturbations, aligning with AI safety concerns.
  verification that quantifies, before deployment, how far a camera-pose perturbation can move the commanded actionprobabilistic reachable-action radius with finite-sample coverage
• Probabilistic Roboticsframework60%
  The approach handles uncertainty in camera poses and provides probabilistic guarantees, aligning with probabilistic robotics.
  camera-pose perturbations are sampled from the prescribed distributionprobabilistic reachable-action radius
• Machine Learningsubfield50%
  The method involves training objectives and conformal prediction, which are machine learning techniques.
  set-based trainingsplit conformal calibration
• Learning-Based Manipulationframework50%
  The experiments are on manipulation tasks, and the method is applied to manipulation policies.
  controlled manipulation experimentsLIBERO-10 manipulation benchmark

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.

Paper Context