What 1,228 Vision-Language-Action Papers Say About the Robot Data Problem

Analysis of 1,228 Vision-Language-Action robotics papers reveals that data, not architecture or compute, is the binding constraint on VLA progress. The field has shifted from needing more data to requiring better-annotated, diverse, and verified datasets, citing three critical gaps: impoverished language annotations, discarded failure data, and missing tactile and force sensing modalities.