Coding Agents Beat Hand-Engineered Task and Motion Planners

Coding agents write task-and-motion programs that beat hand-engineered planners, with 56–95% mean success versus 47% and far less compute.

Researchers tested whether coding agents can synthesize programs for generalized task and motion planning, replacing TAMP-specific engineering. Claude Code using Opus 5 and Codex using GPT-5.6 Sol and GPT-6 Astra were given a task description and simulator access, wrote programs within a fixed budget, and then had those programs frozen for evaluation. Across 28 KinDER and PDDLStream environments, 980 programs were each run on 100 held-out instances, totaling 98,000 episodes. All three agent configurations beat hand-engineered planners, one-shot generation, and an LLM generalized-planning baseline, with mean success of 56% to 95% versus 47% for planners on the 16 environments where a planner exists. As object counts increase, the generated programs remain more successful while using about an order of magnitude less computation per instance. The two reports agree on these results; the daily-papers note also says the authors release all code and the full prompts given to the agents.

• Claude Code (Opus 5) and Codex (GPT-5.6 Sol and GPT-6 Astra) wrote programs from a task description and simulator access, then froze them.
• Tests covered 28 KinDER and PDDLStream environments, 980 programs, and 98,000 episodes (100 held-out instances each).
• Mean success was 56% to 95%, versus 47% for hand-engineered planners on the 16 environments where a planner exists.
• All three agent setups also beat one-shot generation and an LLM generalized-planning baseline.
• As object counts grow, generated programs stay more successful and use about an order of magnitude less computation per instance.
• Programs were synthesized within a fixed budget; the authors release all code and the full prompts.