Teach and Grow: An Agent-Centered Architecture for General Robot Learning

Details

Current end-to-end robot learning (VLA and world-action models) improves with scale but carries a hidden tax: every time the robot encounters an unfamiliar object, environment, or failure mode outside its validated coverage, correcting the failure requires new robot data, a policy update, and regression testing. The authors call this the "retraining tax." Teach-and-Grow Learning (TGL), from Chang Nie, Zhe Liu, and Hesheng Wang at Shanghai Jiao Tong University, proposes an agent-centered architecture to reduce this tax. A multimodal agent converts a few successful demonstrations into Skill Blocks — closed-loop behaviors for meaningful subgoals. In a new scene, the agent grounds and composes available Skill Blocks, selects learned or geometric tools, observes the physical outcome, and revises its plan when execution departs from intent. A Skill Library stores executable behaviors; a structured Experience Memory carries forward successes, failures, and repairs. New tasks are acquired without task-specific policy retraining — the agent adapts by composing and revising existing skills rather than training from scratch. LIBERO evaluation attains state-of-the-art performance; controlled studies demonstrate skill induction, persistent reuse across tasks, and agent-directed adaptation to failure. The authors also propose the "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 — analogous to language model scaling laws but for embodied robot experience.