Teach and Grow: An agent-centric architecture for general robot learning
[Daily Paper] Teach and Grow: An Agent-Centered Architecture for General Robot Learning
AI summary
Chang Nie et al. propose a training-free framework, Teach-and-Grow Learning (TGL), in which a robot can acquire, adapt to, and retain manipulation behaviors without gradient updates, fine-tuning, or weight modification. Targeting the "retraining tax" of end-to-end VLA/WAM, the framework splits a task into a semantic policy that is invariant across environments and a physical grounding computed dynamically from observations, with the base model's parameters frozen. Validated trajectories can be distilled offline into a fast feedforward student policy.