VISTA: Vision-Grounded and Physics-Validated Adaptation of UMI data for VLA Training
Researchers have developed VISTA, a framework designed to improve the training of Vision-Language-Action (VLA) models using real-world robot data. VISTA addresses two key issues: the mismatch between typical fisheye robot camera views and standard VLM representations, and the inclusion of physically infeasible actions in human-collected trajectories. The framework includes a VQA dataset for distorted visual alignment, a pipeline for scoring and filtering trajectories based on physical validity, and a co-training method to learn both grounding and action prediction. AI
IMPACT Enhances VLA model training by addressing data quality and representation mismatches, potentially improving real-world robot deployment.