AquaMend: Minimal Re-probing and Conditional Rollback for Latent-Belief Failures in Embodied Agents

2026-09-23

AI Summary

This study addresses task interruptions in embodied agents caused by belief invalidation due to physical changes or perceptual noise. To tackle this, we propose a Probe-Belief-Action graph modeling framework that evaluates re-probing, rollback, and continuation strategies by minimizing expected loss. The core innovation lies in a conditional detection power screening mechanism guided by the joint posterior, which enables online one-step decision-making with minimal re-probing and conditional rollbacks. Experimental results across 32 test scenarios demonstrate successful recovery in 28 cases. Compared to baseline restart strategies, the proposed approach reduces cumulative loss by 21.6% and decreases online decision latency by 12.3%, thereby validating both its effectiveness and real-time capability.

Abstract

Physical changes or sensing errors can invalidate embodied agents' task-relevant beliefs. AquaMend compares re-probing, rollback, and supported continuation on a probe-belief-action graph under an expected-loss objective covering sensing, physical recovery, and uncorrected failures. A joint posterior guides a one-step policy with conditional detection-power screening. The per-belief three-way optimum requires independence, separability, and fully resolving probes; the general policy has no global optimality guarantee. Across 32 paired scenarios in a self-constructed simulation benchmark, AquaMend recovers in 28/32 cases and reduces mean complete loss by 21.6% versus restart. Its paired loss difference from decision-theoretic troubleshooting (DTT) is not statistically significant after Holm correction. Against the all-candidate ablation, online decision time decreases by 12.3% overall but increases by 3.4% in the uncovered late stage.

Problem

• embodied agents
• belief failures
• sensing errors
• physical changes

Innovation

• Embodied Agents
• Probe-Belief-Action Graph
• Conditional Rollback
• Expected-Loss Objective
• Joint Posterior