Recoleta
GitHub
Source note

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

Authors
Ralf Römer, Maximilian Seeliger, Saida Liu, Ben Sturgis, Marco Bagatella, Daniel Marta, et al.
Source
arXiv
Published
Collection
Embodied AI
Vision Language ActionUncertainty QuantificationFlow MatchingActive Fine TuningRobot Data Scaling
Open arXivMarkdown source

Summary

The paper gives flow-based vision-language-action models a practical confidence signal. It uses ensemble disagreement in the flow velocity field to detect likely failures and to choose which robot demonstrations to collect for fine-tuning.

Problem

  • Flow-based VLAs can act with no reliable warning when a scene, object, or task is outside their training distribution, which matters for real robot deployment and safety.
  • Existing VLA fine-tuning often needs many expert demonstrations, and those demonstrations are expensive to collect.
  • The key missing signal is epistemic uncertainty: uncertainty that should drop when the model receives the right new data.

Approach

  • The method trains a small ensemble of flow-based VLA action heads and compares their predicted velocity fields during action generation.
  • It derives velocity-field disagreement (VFD) as a tractable estimate tied to pairwise KL divergence between flow models, avoiding expensive likelihood computation through ODE divergence terms.
  • In simple terms, the model starts generating an action chunk, checks whether ensemble members point in different action directions along the flow path, and treats larger disagreement as higher uncertainty.
  • The SAVE active fine-tuning method ranks tasks by mean VFD, selects the most uncertain initial observations, asks an expert for demonstrations, and fine-tunes with new data plus replay data.

Results

  • On the LIBERO benchmark, VFD is reported to give better-calibrated uncertainty estimates that predict downstream task performance better than tested baselines; the excerpt does not provide exact calibration or correlation numbers.
  • VFD is reported to perform well for deployment failure detection; the excerpt does not provide AUROC, precision, recall, or threshold numbers.
  • SAVE is reported to need at least 22 fewer samples than baselines for multitask adaptation on LIBERO; the excerpt renders the unit as “22” and does not show whether this means 22 demonstrations or 22%.
  • The method is evaluated on multitask pools with K tasks and L candidate initial observations per task, using R active-learning rounds and n_e expert queries per round, but the excerpt does not give the concrete K, L, R, or n_e values.

Link