Post-hoc Epistemic Uncertainty for Vision-Language-Action Policies
Abstract
Vision-language-action (VLA) policies need to recognize gaps in their training knowledge to request useful demonstrations or assistance in new situations, especially when failures are costly. Disagreement within an ensemble of independently trained policies can guide these decisions, but requires training and running multiple policies. We propose post-hoc kernel-variance heads that estimate epistemic uncertainty from the embeddings of a single policy. Each head computes the conditional variance of a Gaussian process conditioned on embeddings from the training demonstrations. Fitting requires no regression targets or gradient updates and leaves the policy unchanged. In experiments with SmolVLA on LIBERO-10, the kernel variance heads perform competitively with two-policy ensemble scores in task-level correlation with task success, in demonstration acquisition during active learning, and in referral to a stronger supervisor policy as a proxy for human assistance. For SmolVLA, the head on the vision-language backbone reduces uncertainty-scoring time by up to approximately a factor of 50 compared with ensemble scores, while also substantially reducing setup costs. We also show that the established evaluation protocol on LIBERO-10 primarily rewards recognition of coverage differences between tasks, while execution variability weakens assessment of within-task rankings. Thus, controlled perturbations provide a complementary evaluation: modality-specific heads identify unfamiliar camera, instruction, or state inputs, while fused heads can detect unfamiliar combinations of individually familiar inputs. With a supervisor proxy trained on the evaluated perturbation families, the decoder head improves within-task referral over ensemble disagreement. These findings show that useful uncertainty estimates for adaptive robot learning can be obtained from a single policy.
Related
- A Three-Way Testing Framework for Quantifying Epistemic Calibration Uncertainty in SBI
- Hybrid Uncertainty Sensitivity Analysis Based on the HSIC for High-Dimensional Responses with Aleatory--Epistemic Separation
- Inventory Policies with Returns and Demand Uncertainty: Optimal, Fixed-Quantity, and Order Cancellation Models
- What to Reuse in Generative Robot Policies? A Solver-Aware Mathematical Analysis of Training-Free Output Caching for Action Generation
- Sampling Uncertainty is Critical for Setting Forecast-Based Anticipatory Action Triggers
BibTeX
@misc{keller2026posthoc,
title = {{Post-hoc Epistemic Uncertainty for Vision-Language-Action Policies}},
author = {Johannes Keller and Albrecht M. Bloße and Adrian Lindenmeyer and Thomas Neumuth and Daniel Schneider},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.29730},
eprint = {2610.29730},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.29730},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Post-hoc Epistemic Uncertainty for Vision-Language-Action Policies
AU - Johannes Keller
AU - Albrecht M. Bloße
AU - Adrian Lindenmeyer
AU - Thomas Neumuth
AU - Daniel Schneider
PY - 2026
DA - 2026/10/21
PB - arXiv
DO - 10.48550/arxiv.2610.29730
UR - https://arxiv.org/abs/2610.29730
AB - Vision-language-action (VLA) policies need to recognize gaps in their training knowledge to request useful demonstrations or assistance in new situations, especially when failures are costly. Disagreement within an ensemble of independently trained policies can guide these decisions, but requires training and running multiple policies. We propose post-hoc kernel-variance heads that estimate epistemic uncertainty from the embeddings of a single policy. Each head computes the conditional variance of a Gaussian process conditioned on embeddings from the training demonstrations. Fitting requires no regression targets or gradient updates and leaves the policy unchanged. In experiments with SmolVLA on LIBERO-10, the kernel variance heads perform competitively with two-policy ensemble scores in task-level correlation with task success, in demonstration acquisition during active learning, and in referral to a stronger supervisor policy as a proxy for human assistance. For SmolVLA, the head on the vision-language backbone reduces uncertainty-scoring time by up to approximately a factor of 50 compared with ensemble scores, while also substantially reducing setup costs. We also show that the established evaluation protocol on LIBERO-10 primarily rewards recognition of coverage differences between tasks, while execution variability weakens assessment of within-task rankings. Thus, controlled perturbations provide a complementary evaluation: modality-specific heads identify unfamiliar camera, instruction, or state inputs, while fused heads can detect unfamiliar combinations of individually familiar inputs. With a supervisor proxy trained on the evaluated perturbation families, the decoder head improves within-task referral over ensemble disagreement. These findings show that useful uncertainty estimates for adaptive robot learning can be obtained from a single policy.
ER -