VANE: Reliable Test-Time Training for Vision-Language-Action Models via Future Visual Representation Prediction

2026-08-10

AI Summary

Existing test-time training methods in closed-loop vision–language–action control often suffer from task interference and action perturbations due to shared adaptation spaces. This work proposes the VANE framework, which introduces a context-conditioned prompt adaptation mechanism coupled with prediction of future visual representations to enable decoupled, reversible updates relative to the online policy: candidate adaptations are committed only when subsequent observations provide supporting evidence. This constrained, evidence-driven approach ensures reliable and task-consistent test-time adaptation. Experiments demonstrate that VANE improves average success rates by 3.2 percentage points on SimplerEnv WidowX; deployment gains on the Google Robot exhibit notable task- and embodiment-dependence.

Abstract

Test-time training (TTT) offers a lightweight way to adapt vision--language--action (VLA) policies from unlabeled deployment streams, but it remains difficult to use reliably in closed-loop manipulation. A shared adaptation space can mix incompatible task corrections, while an online update can alter subsequent actions before its consequences are known. We introduce a reliable TTT framework for VLA policies (VANE). VANE conditions prompt adaptation on the current vision--language context and learns from the future visual consequences of executed actions. Candidate updates are isolated from the live policy, evaluated on subsequent observations, and committed only when supported by future evidence, making adaptation selective and reversible. On SimplerEnv WidowX, VANE improves average success by $3.2$ percentage points over the corresponding TTT baseline. Results on Google Robot further show that deployment-time gains remain task- and embodiment-dependent. Together, these results demonstrate a constrained, evidence-based approach to adapting VLA policies during interaction.

Problem

• test-time training
• vision-language-action models
• closed-loop manipulation
• policy adaptation
• reliable adaptation

Innovation

• test-time training
• vision-language-action models
• future visual prediction
• selective adaptation
• evidence-based policy update