Learning to Move Before Learning to Do: Task-Agnostic pretraining for VLAs

Vision-Language-Action (VLA) models for robotics are bottlenecked by scarce, costly expert demonstrations that combine observations, instructions, and actions together. This paper argues that physical competence (how to move) and semantic task understanding (what to do) can be learned separately, with only the latter needing costly labeled data. Their proposed Task-Agnostic Pretraining (TAP) first learns motor skills from cheap, unlabeled interaction data, then grounds these skills in language using minimal expert examples. TAP matches models trained on over a million demonstrations while using far less labeled data and remains far more robust to real-world visual perturbations like camera changes. Authors: Junhao Shi, Siyin Wang, Xiaopeng Yu, Li Ji, Jingjing Gong, Xipeng Qiu Paper: https://arxiv.org/abs/2607.02466v1