Coding Agents for Generalized Task and Motion Planning Problems
Coding agents beat available hand-engineered TAMP planners on mean success while using far less per-instance compute.
The paper tests Claude Code and Codex variants on 28 simulated robotics environments from KinDER and PDDLStream. The agents synthesize programs under a fixed budget, then those programs are frozen and run on unseen instances. Across 98,000 evaluation episodes, the agent-written programs reached 56% to 95% mean success on environments with planner baselines, versus 47% for the planners. The authors say the agents’ logs show iterative testing: calibrating physical models, checking edge cases, and revising strategies. HF Daily Papers' note
The paper tests Claude Code and Codex variants on 28 simulated robotics environments from KinDER and PDDLStream. The agents synthesize programs under a fixed budget, then those programs are frozen and run on unseen instances. Across 98,000 evaluation episodes, the agent-written programs reached 56% to 95% mean success on environments with planner baselines, versus 47% for the planners. The authors say the agents’ logs show iterative testing: calibrating physical models, checking edge cases, and revising strategies. HF Daily Papers' note
score 5