Teach-and-Grow robotics architecture eliminates retraining tax with reusable Skill Blocks

A team from researchers Chang Nie, Zhe Liu, and Hesheng Wang has introduced Teach-and-Grow Learning (TGL), a new agent-centered architecture designed to overcome the 'retraining tax' that plagues end-to-end vision-language-action (VLA) and world-action models. These models fail when encountering unfamiliar objects, sensors, or contact scenarios, requiring new robot data, policy updates, and regression testing. TGL instead lets a multimodal agent turn a few successful demonstrations into reusable Skill Blocks—closed-loop behaviors tied to meaningful subgoals. In new scenes, the agent grounds and composes these blocks, selects learned or geometric tools, observes physical outcomes, and revises routes when execution strays from intent.

The architecture relies on a Skill Library to store executable behavior and structured Experience Memory to carry forward success, failure, and repair records. This enables new tasks to be acquired without task-specific policy retraining. In LIBERO benchmark evaluations, TGL achieves state-of-the-art performance, with controlled studies demonstrating skill induction, persistent reuse, and agent-directed adaptation. The authors also propose a Teach-and-Grow scaling-law hypothesis: if X denotes effective reusable experience, future-task error and teaching demand should approach irreducible floors as power laws in X. This positions deployment as a period of continued learning, where one task makes the next easier—paving the way for robots that truly learn and grow over time.