Robotics · 2026-10-26

Do VLAs Know They Will Fail? Interpretable Success and Failure Patterns in Physical AI Systems

Christian A. Rogers, Qian Lin, Daniel S. Brown

arXiv:2610.32208PDF

AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.

TL;DR

The authors show that sparse autoencoder (SAE) representations of OpenVLA activations contain outcome-predictive information and can predict task failure before an estimated point of no return. They identify individual SAE features whose strong activations correlate with success or failure and localize critical temporal intervals where features rise. A VLM-based pipeline generates behavioral descriptions of these patterns, validated via detection scoring on held-out clips.

Why it matters

Diagnosing why vision-language-action (VLA) rollouts fail is difficult from dense internal representations. By decomposing OpenVLA activations into sparse SAE features, the authors connect early failure warning to interpretable, outcome-associated behavioral patterns, and validate that VLM-generated descriptions align with feature activations on held-out data.

Method

  • Train logistic regressors at each timestep on rolling-window (T=40) raw OpenVLA activations or rolling-mean SAE features to predict eventual failure; compare against a majority baseline.
  • Define a simulator-based point of no return (PNR) using estimated recoverability from restored states and stochastic continuations; trigger alarms when predicted failure probability exceeds a threshold for two consecutive evaluations.
  • Select SAE features by success rate over top-k trajectories, localize one critical “rising” activation interval per trajectory, then use a VLM to generate behavioral descriptions and validate them with AUROC detection scoring on held-out clips.

Limitation

Recoverability estimation is an imperfect instrument: remaining step budget and temperature sampling can affect whether branches succeed, late PNRs are excluded after step 420, and the resulting recoverability measures recoverability for the policy rather than physical recoverability.

Abstract (from arXiv)

Vision-Language-Action (VLA) models have emerged as powerful general-purpose robot policies, but their successes and failures remain difficult to diagnose and understand. In this work, we show that sparse autoencoder (SAE) representations of OpenVLA activations contain outcome-predictive information and can be used to predict task failure before the robot reaches an unrecoverable state. Moreover, to provide insights into what common behaviors may contribute to these outcomes, we identify individual SAE features whose strong activations are associated with specific stages of success or failure, and uncover recurring behavioral patterns by localizing temporal intervals where these features exhibit significant activation changes. We further develop a VLM-based interpretation pipeline to generate behavioral descriptions of these patterns and validate their correspondence to feature activations on held-out intervals.

Related papers

Publication record

Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.

First public release: 2026-10-23 · Indexed by Pub-AI: 2026-10-26 UTC

preprint

Do VLAs Know They Will Fail? Interpretable Success and Failure Patterns in Physical AI Systems

Christian A. Rogers, Qian Lin, Daniel S. Brown

Preprint posted: 2026-10-23 · Original record ↗

arxiv
2610.32208
Explicit source versions (1)
  • v1 · 2026-10-23
  1. Preprint available

    Source date: 2026-10-23 · First detected: 2026-11-07 UTC · Historical baseline

    Source evidence (1)
Abstract · arxiv

Vision-Language-Action (VLA) models have emerged as powerful general-purpose robot policies, but their successes and failures remain difficult to diagnose and understand. In this work, we show that sparse autoencoder (SAE) representations of OpenVLA activations contain outcome-predictive information and can be used to predict task failure before the robot reaches an unrecoverable state. Moreover, to provide insights into what common behaviors may contribute to these outcomes, we identify individual SAE features whose strong activations are associated with specific stages of success or failure, and uncover recurring behavioral patterns by localizing temporal intervals where these features exhibit significant activation changes. We further develop a VLM-based interpretation pipeline to generate behavioral descriptions of these patterns and validate their correspondence to feature activations on held-out intervals.