Teach-and-Grow Learning: New Agent-Centered Architecture for General Robot Learning
A recent preprint on arXiv (2608.17209) introduces Teach-and-Grow Learning (TGL), an architecture focused on enhancing robot learning from the agent's perspective. It addresses the 'retraining tax' associated with end-to-end vision-language-action (VLA) and world-action models, which face limitations due to their validated physical coverage. When robots encounter unfamiliar objects, sensors, embodiments, or contacts outside this coverage, rectifying issues necessitates new data, policy revisions, and regression testing. Unlike textual data, embodied data often requires machine operation. TGL transforms successful demonstrations into reusable Skill Blocks—closed-loop behaviors aimed at achieving specific subgoals. In new environments, the agent utilizes these blocks, chooses learned or geometric tools, monitors physical results, and adjusts paths if execution diverges. This preprint is designated as a cross announcement.
Key facts
- The arXiv preprint 2608.17209 introduces Teach-and-Grow Learning (TGL).
- TGL is an agent-centered architecture for general robot learning.
- It addresses the 'retraining tax' in VLA and world-action models.
- The reliability of end-to-end VLA and world-action models is bounded by validated physical coverage.
- Failures outside coverage require new robot data, policy updates, and regression testing.
- Embodied data must often be created by operating machines, unlike text.
- TGL turns a few successful demonstrations into reusable Skill Blocks.
- In new scenes, the agent grounds and composes Skill Blocks, selects tools, observes outcomes, and revises routes.
Entities
—