World In-Context Learning for Agentic Robot Manipulation
Trained action policies write manipulation skill into network weights, tying what is learned to the distribution it is learned from. In-context learning (ICL) avoids this coupling but has not matched
- Upvotes11