Coding Agents for Generalized Task and Motion Planning Problems

Coding Agents for Generalized Task and Motion Planning Problems. By Matteo Merler and six others, from Princeton University, Fondazione Bruno Kessler, and two other institutions. Published on the twenty-fourth of September. Its code repository is worked out: about two hundred and fifty source files, with a readme and a licence.

Task and motion planning, or TAMP, stays hard even when everything is observable and the state is described object by object. The reason is coupling. The discrete decisions of a plan are tightly bound to geometric, kinematic, and dynamic constraints. Generalized TAMP eases this by exploiting regularities across problem instances, so that planning a new instance takes less effort. The catch is that existing methods demand a great deal of engineering specific to TAMP. The authors ask whether coding agents can automate that engineering, by writing programs that generalize across instances. Their answer is yes, in the authors' own word surprisingly so: coding agents are surprisingly effective at generalized TAMP.

The setup is strict about separation. Each agent receives a task description and access to a simulator. It decides for itself how to interact with the environment while it develops a program, inside a fixed synthesis budget. Then the program is frozen and evaluated on instances it has never seen. The agents are Claude Code, running Opus 5, and Codex, running two models, GPT-5.6 Sol and GPT-6 Astra. They are tested on twenty-eight simulated environments drawn from KinDER and PDDLStream, with object counts beyond those in the original benchmark.

The scale is large. Across every program synthesis method, nine hundred and eighty generated programs are each evaluated on one hundred held-out instances, for ninety-eight thousand evaluation episodes in total. All three agent configurations beat hand-engineered planners, one-shot generation, and a generalized planning baseline based on a language model, on mean success. On the sixteen environments where a planner exists, the agents reach between fifty-six and ninety-five percent success. The planners reach forty-seven percent. As object counts grow, the agents' programs keep a higher success rate than the planner, while using on average an order of magnitude less computation per instance.

The logs show how the agents used their access. They interacted with the simulator to calibrate physical models, to test edge cases, and to refine their strategies before the program was frozen.

The claim has clear edges. It is about simulated environments, drawn from two named suites, and each program is judged after a fixed synthesis budget has been spent. The comparison with planners rests on the sixteen environments where a planner is available, not on all twenty-eight. And the authors state their conclusion with matching care. They suggest that coding agents are a strong baseline for generalized TAMP: a reference point that other approaches can be measured against. All the code is released, including the full prompts the agents were given.

Engineering that generalized TAMP has required from people is recast here as a programming task, handed to an agent with a simulator and a budget. On these environments the handover works. The programs the agents write generalize to unseen and larger instances, beat the hand-built planners, and run cheaper while doing it.