aiqbench analysis
This research addresses a critical gap in Vision-Language-Action models by introducing a training-free method to quantify epistemic uncertainty via weight perturbations. By leveraging disagreement across perturbed predictions, the framework effectively detects failures under distribution shifts where traditional stochastic sampling falls short. The results demonstrate significant improvements in failure detection metrics for robotic manipulation tasks facing unseen object variations.
Perturbation-Based Epistemic Uncertainty for Failure Detection in Vision-Language-Action Models
arXiv:2606.20754v2 Announce Type: replace Abstract: Vision-Language-Action (VLA) models have shown strong performance in robotic manipulation, but reliable uncertainty quantification remains challenging, particularly under distribution shift. Unlike autoregressive policies, many modern VLA models generate continuous actions through regression or flow-based generation, where explicit predictive probabilities are unavailable. Moreover, stochastic action sampling primarily captures action-generation variability under a fixed model, while failure detection under distribution shift can benefit from capturing uncertainty in the model itself. Motivated by Bayesian perspectives on local model variations, we propose perturbation-based failure detection (PFD), a training-free framework for estimating epistemic uncertainty in VLA models through low-rank weight perturbations. Specifically, we inject random low-rank perturbations into selected transformer weight matrices and estimate epistemic uncertainty from disagreement across perturbed action predictions. Experiments on LIBERO-PRO show that PFD achieves the highest average AUROC and balanced accuracy among the evaluated methods while consistently outperforming stochastic action sampling across distribution shifts. Real-world robot experiments further demonstrate that PFD provides a competitive failure-detection signal under an unseen object shift.
Abstract:Vision-Language-Action (VLA) models have shown strong performance in robotic manipulation, but reliable uncertainty quantification remains challenging, particularly under distribution shift. Unlike autoregressive policies, many modern VLA models generate continuous actions through regression or flow-based generation, where explicit predictive probabilities are unavailable. Moreover, stochastic action sampling primarily captures action-generation variability under a fixed model, while failure detection under distribution shift can benefit from capturing uncertainty in the model itself. Motivated by Bayesian perspectives on local model variations, we propose perturbation-based failure detection (PFD), a training-free framework for estimating epistemic uncertainty in VLA models through low-rank weight perturbations. Specifically, we inject random low-rank perturbations into selected transformer weight matrices and estimate epistemic uncertainty from disagreement across perturbed action predictions. Experiments on LIBERO-PRO show that PFD achieves the highest average AUROC and balanced accuracy among the evaluated methods while consistently outperforming stochastic action sampling across distribution shifts. Real-world robot experiments further demonstrate that PFD provides a competitive failure-detection signal under an unseen object shift.
Robotics (cs.RO)
arXiv:2606.20754 [cs.RO]
(or arXiv:2606.20754v2 [cs.RO] for this version)
https://doi.org/10.48550/arXiv.2606.20754 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 252 words
Related coverage
Uncertainty Quantification for Flow-Based Vision-Language-Action Models
arXiv · Jun 17, 2026
Uncertainty Quantification for Flow-Based Generalist Robot Policies
arXiv · Oct 5, 2026
You Are What You Prompt: Prompt Quality, Domain Shift, and Uncertainty in Agrifood Vision-Language Models
arXiv · Aug 20, 2026
Coverage-driven alignment - What ‘Teaching Claude Why’ can borrow from AV verification
LessWrong · Jun 8, 2026