Embodied Interpretability: Linking Causal Understanding to Generalization in Vision-Language-Action Models

Fuente: arXiv
Guardado en:
Detalles Bibliográficos
Autores principales: Zhang, Hanxin, Xu, Mingshuo, Dhafer, Abdulqader, Yue, Shigang, Dong, Hongbiao, Hao, Zhou Daniel
Formato: Preprint
Publicado: 2026
Materias:
Acceso en línea:
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866914524123103232
author Zhang, Hanxin
Xu, Mingshuo
Dhafer, Abdulqader
Yue, Shigang
Dong, Hongbiao
Hao, Zhou Daniel
author_facet Zhang, Hanxin
Xu, Mingshuo
Dhafer, Abdulqader
Yue, Shigang
Dong, Hongbiao
Hao, Zhou Daniel
contents Vision-Language-Action (VLA) policies often fail under distribution shift, suggesting that decisions may depend on spurious visual correlations rather than task-relevant causes. We formulate visual-action attribution as an interventional estimation problem. Accordingly, we introduce the Interventional Significance Score (ISS), an interventional masking procedure for estimating the causal influence of visual regions on action predictions, and the Nuisance Mass Ratio (NMR), a scalar measure of attribution to task-irrelevant features. We analyze the statistical properties of ISS and show that it admits unbiased estimation, and we characterize conditions under which action prediction error provides a valid proxy for causal influence. Experiments across diverse manipulation tasks indicate that NMR predicts generalization behavior and that ISS yields more faithful explanations than existing interpretability methods. These results suggest that interventional attribution provides a simple diagnostic approach for identifying causal misalignment in embodied policies.
format Preprint
id arxiv_https___arxiv_org_abs_2605_00321
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Embodied Interpretability: Linking Causal Understanding to Generalization in Vision-Language-Action Models
Zhang, Hanxin
Xu, Mingshuo
Dhafer, Abdulqader
Yue, Shigang
Dong, Hongbiao
Hao, Zhou Daniel
Robotics
Vision-Language-Action (VLA) policies often fail under distribution shift, suggesting that decisions may depend on spurious visual correlations rather than task-relevant causes. We formulate visual-action attribution as an interventional estimation problem. Accordingly, we introduce the Interventional Significance Score (ISS), an interventional masking procedure for estimating the causal influence of visual regions on action predictions, and the Nuisance Mass Ratio (NMR), a scalar measure of attribution to task-irrelevant features. We analyze the statistical properties of ISS and show that it admits unbiased estimation, and we characterize conditions under which action prediction error provides a valid proxy for causal influence. Experiments across diverse manipulation tasks indicate that NMR predicts generalization behavior and that ISS yields more faithful explanations than existing interpretability methods. These results suggest that interventional attribution provides a simple diagnostic approach for identifying causal misalignment in embodied policies.
title Embodied Interpretability: Linking Causal Understanding to Generalization in Vision-Language-Action Models
topic Robotics
url https://arxiv.org/abs/2605.00321