Robotics Research Hub
Uncertainty Quantification for Flow-Based Vision-Language-Action Models

Uncertainty Quantification for Flow-Based Vision-Language-Action Models

Featured

Ralf Römer, Maximilian Seeliger, Saida Liu, Ben Sturgis, Marco Bagatella, Daniel Marta, Andreas Krause, Angela P. Schoellig · ETH Zürich / University of Toronto (Learning Systems & Robotics Lab) · 2026

vla foundation-models safety manipulation
PDF arXiv Code (GitHub)

Framework

PyTorch

License

N/A

Stars

N/A

Summary

Quantifies epistemic uncertainty in flow-matching VLAs using velocity-field disagreement (VFD) across a small ensemble, enabling failure detection at deployment and sample-efficient active fine-tuning (SAVE).

Abstract Summary

Vision-language-action models (VLAs) combine vision-language backbones with expressive generative action heads trained via flow matching on large-scale robotic datasets. Despite their strong empirical performance in robotic manipulation, VLAs lack mechanisms to quantify confidence in their predictions and to detect when their actions may be unreliable. This presents a critical limitation for real-world deployment in non-stationary environments, where models inevitably encounter scenarios outside their pretraining distribution and may fail without warning. To address this, the authors derive an efficient method for quantifying epistemic uncertainty in flow-matching models by leveraging velocity-field disagreement (VFD) across a small ensemble. They use this uncertainty estimate for failure detection during deployment and active fine-tuning of flow-based VLAs. They propose SAVE, a framework for uncertainty-guided active multitask fine-tuning that reduces the number of costly expert demonstrations required to adapt VLAs to new tasks. Through extensive experiments on the LIBERO benchmark, they demonstrate that VFD yields better-calibrated uncertainty estimates predictive of downstream performance, that VFD achieves strong performance in detecting failures, and that uncertainty-guided data acquisition with SAVE requires at least 22% fewer samples than baselines.

Key Points

  • Introduces velocity-field disagreement (VFD) for epistemic uncertainty quantification in flow-matching VLAs
  • Uses uncertainty for failure detection during deployment
  • Proposes SAVE: uncertainty-guided active multitask fine-tuning requiring ≥22% fewer expert demonstrations
  • Validated on LIBERO benchmark with better-calibrated uncertainty than baselines
  • RSS 2026 Best Paper Award

Preliminaries

Given a dataset ${\mathcal{D}=\{(\bm{x}^{(n)},\bm{y}^{(n)})\}_{n=1}^{N}}$ and a neural network parameterized by $\bm{\theta}\in\Theta$, trained to learn the relationship between $\bm{y}\in\mathcal{Y}$ and $\bm{x}\in\mathcal{X}$, Bayesian modeling considers the posterior predictive distribution where $p(\bm{\theta}\,/\,\mathcal{D})$ is the posterior and generally intractable for deep neural networks ( Hüllermeier and Waegeman, 2021 ) .

Preliminaries — Uncertainty Quantification for Flow-Based Vision-Language-Action Models
Figure 2: Real-world tasks.

Estimating Epistemic Uncertainty of Flow-Matching Models

Training an FM model yields an estimate $\bm{\theta}$ of the model parameters given the data $\mathcal{D}$. For a new conditioning input $\bm{y}$, how can we quantify whether the model exhibits high epistemic uncertainty that could be reduced by collecting more data? A simple estimate of total predictive uncertainty is the conditional entropy $H[p(\bm{x}\mid\bm{y},\bm{\theta})]$.

However, as this quantity conflates epistemic and aleatoric uncertainty, we need to isolate the epistemic component ( 2 ).

Estimating Epistemic Uncertainty of Flow-Matching Models — Uncertainty Quantification for Flow-Based Vision-Language-Action Models
Figure 3: Calibration w.r.t. the language prompt. VFD is well-calibrated when injecting uncertainty only through this modality.

Uncertainty-Guided Active Multitask Fine-Tuning

Given a flow-based generalist policy and a set of tasks, how can we improve overall performance across tasks with minimum data collection effort?

More formally, we consider a set of $K$ tasks ${\mathcal{T}=\{T_{k}\}_{k=1}^{K}}$, each of which can be described by an MDP ${\mathcal{M}_{k}=(\mathcal{O},\mathcal{A},P_{k},R_{k},\rho_{k},\gamma)}$, where ${P_{k}:\mathcal{O}\times\mathcal{A}\to\Delta(\mathcal{O})}$, ${R_{k}:\mathcal{O}\to\mathbb{R}}$ and ${\rho_{k}\in\Delta(\mathcal{O})}$ are task-specific dynamics, rewards and initial observation dis...

Uncertainty-Guided Active Multitask Fine-Tuning — Uncertainty Quantification for Flow-Based Vision-Language-Action Models
Figure 4: Active fine-tuning in the real world. For fixed demonstration budgets, SAVE w/ VFD improves multitask performance the most.

Experiments

Our experiments in simulation and the real world primarily aim at answering three research questions: How well-calibrated is VFD, i.e., to what extent do its values reflect the epistemic uncertainty and task-completion capabilities of flow-based generalist policies? Can VFD identify where additional data should be collected for active fine-tuning?

How informative are VFD uncertainty estimates for detecting failures during deployment? We conduct our experiments on three benchmarks.

Experiments — Uncertainty Quantification for Flow-Based Vision-Language-Action Models
Figure 5: Exploration vs. exploitation. SAVE w/ VFD prioritizes tasks with higher uncertainty, trading off diversity for multitask improvement.

Sources and Demonstrations

Figures are reproduced from Römer et al. See the full paper for experimental details and the project page, when the paper links one, for demonstrations and videos.

Share

Related Papers