TAP: Task-Agnostic Pretraining for Vision-Language-Action models

Researchers propose Task-Agnostic Pretraining (TAP), a two-stage framework designed to address the scarcity of expert demonstrations in Vision-Language-Action (VLA) models by decoupling physical competence from semantic alignment. The method first learns transferable motor priors from unlabeled interaction data via a self-supervised Inverse Dynamics objective, then grounds these priors in language using minimal expert data.

• TAP matches the performance of models trained on over 1M expert trajectories while using orders of magnitude less labeled data.
• The approach yields a 10% absolute gain over standard behavior cloning on the SIMPLER benchmark.
• On a real-world WidowX platform, TAP retains 25% success under camera perturbations where internet-scale baselines collapse to 0%.

This demonstrates that task-agnostic pretraining produces robust, transferable physical representations and offers a scalable path forward for Embodied AI.