DreamAvoid: Critical-Phase Test-Time Dreaming to Avoid Failures in VLA Policies

2026-05-12

AI Summary

This work addresses the vulnerability of existing Vision-Language-Action models to irreversible failures during fine-grained manipulation, often caused by minor action errors and a lack of explicit awareness of critical states. To mitigate this, the authors propose a test-time “dreaming” mechanism: a Dream Trigger identifies critical phases, an Action Proposer generates candidate action chunks, and a Dream Evaluator—trained on hybrid data—predicts near-future outcomes and assesses action values to select the optimal choice. By integrating dream-based reasoning during critical stages with autonomous boundary learning, the model proactively avoids failure. Experiments demonstrate significant improvements in task success rates on both real-world manipulation tasks and simulation benchmarks, effectively reducing irreversible failures in critical phases.

Abstract

Vision-Language-Action (VLA) models are often brittle in fine-grained manipulation, where minor action errors during the critical phases can rapidly escalate into irrecoverable failures. Since existing VLA models rely predominantly on successful demonstrations for training, they lack an explicit awareness of failure during these critical phases. To address this, we propose DreamAvoid, a critical-phase test-time dreaming framework that enables VLA models to anticipate and avoid failures. We also introduce an autonomous boundary learning paradigm to refine the system's understanding of the subtle boundary between success and failure. Specifically, we (1) utilize a Dream Trigger to determine whether the execution has entered a critical phase, (2) sample multiple candidate action chunks from the VLA via an Action Proposer, and (3) employ a Dream Evaluator, jointly trained on mixed data (success, failure, and boundary cases), to "dream" the short-horizon futures corresponding to the candidate actions, evaluate their values, and select the optimal action. We conduct extensive evaluations on real-world manipulation tasks and simulation benchmarks. The results demonstrate that DreamAvoid can effectively avoid failures, thereby improving the overall task success rate. Our code is available at https://github.com/XianzheFan/DreamAvoid.

Problem

• Vision-Language-Action
• failure avoidance
• critical phase
• manipulation robustness
• test-time adaptation

Innovation

• test-time dreaming
• critical-phase detection
• failure anticipation
• vision-language-action models
• autonomous boundary learning