VISTA: Vision-Grounded and Physics-Validated Adaptation of UMI data for VLA Training

VISTA is a new framework that improves robot learning by adapting real-world manipulation data collected via Universal Manipulation Interface (UMI) for training Vision-Language-Action (VLA) models. The framework addresses two key challenges: making distorted wrist-mounted camera views compatible with pre-trained vision models and filtering out physically infeasible trajectories before training, resulting in significantly better policy performance.

VISTA represents a meaningful advancement in robotics and embodied AI by tackling practical data quality challenges that have limited the scaling of vision-language-action models. The framework identifies and solves two distinct but interconnected problems in real-world robot learning: the domain mismatch between specialized robotic hardware (wrist-mounted fisheye cameras with radial distortion) and general-purpose pre-trained vision models, and the reality that human-collected demonstrations often contain physically impossible or unsafe actions. The introduction of UMI-VQA, a vision-question-answering dataset specifically designed for distorted fisheye perspectives, bridges the visual grounding gap through targeted auxiliary supervision. The physical-validation pipeline then functions as a quality gate, scoring trajectories across multiple dimensions including collision risk and controller bandwidth constraints. This dual-stage approach reflects a growing sophistication in how researchers approach sim-to-real transfer and real-world data utilization. For the broader robotics community, VISTA's public release of validated datasets, the validation pipeline, and pre-trained models accelerates progress by providing infrastructure for downstream researchers. The empirical validation across both simulation and real tasks demonstrates that the framework generalizes beyond single-domain testing. This work signals that success in scaling robotic learning depends not just on collecting more data but on developing systematic methods to curate and adapt that data. The framework's ability to consistently outperform established baselines suggests that quality curation methods may be as important as raw data volume in developing deployable robotic systems.

• →VISTA addresses domain mismatch between distorted wrist-camera views and pre-trained vision models through auxiliary vision-language supervision on specialized fisheye data.
• →A systematic physical-validation pipeline filters out kinematically infeasible trajectories before training, improving deployment success rates.
• →Co-training on curated datasets substantially improves Vision-Language-Action model performance compared to strong existing baselines.
• →Open release of validation tools, UMI-VQA dataset, and pre-trained models reduces barriers for robotics researchers implementing similar approaches.
• →Physical-validation scores predict real-world deployment success, enabling data-centric approaches to robotic policy learning.