Summary

VISTA tackles two mismatches when using Universal Manipulation Interface (UMI) handheld-gripper data to train VLA models: wrist-mounted fisheye views are visually out-of-distribution for pretrained VLMs, and human-collected UMI trajectories often violate robot kinematic limits or exceed controller bandwidth, teaching policies physically infeasible actions. It introduces UMI-VQA, an 8M-pair vision-language dataset built for fisheye wrist views, a trajectory-level physical-validation pipeline that filters infeasible demonstrations before training, and a two-stage co-training recipe (VQA-action alignment, then a flow-matching action expert).

Key Contributions

  • UMI-VQA: a large-scale (8M-pair) VQA dataset tailored specifically to wrist-fisheye observations.
  • An automated physical-feasibility validation and curation pipeline for UMI-style human data.
  • A two-stage co-training recipe integrating perceptual alignment with flow-matching action generation.

Strengths

  • Directly targets a known scalability bottleneck: using cheap handheld UMI data safely for VLA training.
  • The physical-validation step is a concrete, reusable data-quality contribution beyond this specific paper.
  • Open-source code released alongside the paper.

Weaknesses

  • Effectiveness is tied to UMI-style data collection and may not generalize to other teleoperation modalities.
  • The false-negative/positive filtering rate of the physical-validation pipeline is not clearly reported.
  • The two-stage co-training recipe adds pipeline complexity relative to simple end-to-end fine-tuning.

Open Questions

  • How much of the improvement comes from UMI-VQA versus the physical-validation filtering, in isolation?
  • Does the fisheye-specific VQA transfer to standard third-person-camera VLA training?
  • What fraction of raw UMI demonstrations does the validation pipeline discard?

Significance

A concrete, reusable data-quality pipeline for one of the cheapest available manipulation data-collection modalities (handheld UMI grippers), directly matching the cross-cutting data-augmentation priority for VLA training.