What's this about?
Claude Code and Codex interact with simulators, infer physics and edge cases, and write programs that are frozen before evaluation on larger unseen task-and-motion instances.
What exactly is task-and-motion planning?
A robot must choose both what to do and physically how to do it: “put the cup away” includes selecting an order, finding a collision-free path and respecting reach and stability constraints.
Why does this matter?
author-reported preprint; 11 HF upvotes.
- The study evaluates 980 generated programs over 98,000 held-out episodes.
- Agent-written planners outperform available hand-engineered planners in mean success and use roughly one-tenth the per-instance compute.
- Results come from 28 simulated environments; transfer to noisy physical robots is not established.