Confidence-Gated Robot Autonomy: When Does Uncertainty Actually Help?

2026-05-18

AI Summary

This study investigates how predictive uncertainty can guide robotic agents to autonomously execute or switch to fallback strategies in temporal activity recognition tasks, particularly under distribution shifts. To this end, the authors propose a decision-impact-oriented uncertainty evaluation framework that systematically assesses the gating efficacy of various uncertainty quantification methods across multiple benchmarks and temporally embodied simulation environments, leveraging metrics such as Spearman rank correlation, paired bootstrap equivalence tests, and execution/handoff consistency. The findings reveal that when model capability exceeds a certain threshold, diverse uncertainty methods converge in performance, with threshold selection dominating gating outcomes; below this threshold, uncertainty rankings become unstable. Furthermore, while ranking robustness persists under temporal covariate shift, fine-grained semantic anomaly detection performs near-randomly, highlighting a fundamental distinction between gating control and novelty detection.

Abstract

Robotic systems often use predictive uncertainty to decide whether to act autonomously or defer to a fallback policy. In threshold-gated autonomy, uncertainty matters mainly through its ability to rank likely errors. Standard metrics such as expected calibration error and AUROC do not directly test whether uncertainty changes act/defer decisions. We therefore evaluate uncertainty using Spearman rank correlation, paired bootstrap equivalence testing, and act/defer agreement. Across three temporal activity-recognition benchmarks, we find a dataset-dependent competence regime below which uncertainty provides a weak and unstable error ranking. Above this regime, softmax heuristics, MC Dropout, and ensembles produce similar gating behavior, while threshold choice has a much larger effect on execution outcomes. A multi-seed embodied simulation shows the same pattern for collision rate and cost once realized autonomy is matched. Under temporal covariate shift, ranking quality remains stable, but fine grained semantic OOD detection remains near chance. These results suggest that simple uncertainty proxies can suffice for selective gating once the base model is competent, but not for semantic novelty detection.

Problem

• robot autonomy
• predictive uncertainty
• act/defer decision
• temporal activity recognition
• semantic OOD detection

Innovation

• uncertainty quantification
• robot autonomy
• selective gating
• temporal activity recognition
• out-of-distribution detection