Programming agents break into robot motion, with code synthesis outperforming manual plans in generalized planning tests

Task and motion planning in robotics has remained one of the most complex programming challenges, as separate logical decisions intertwine intricately with the geometric, kinematic and dynamic constraints of the real world. Although attempts have been made to develop generalized planning frameworks that can transfer expertise across different situations to reduce computational effort, those systems have continued to require intensive, costly manual engineering to craft each solution individually. A new research paper presented by a joint team including Matteo Merler, Boin Lee, Tom Silver and colleagues reveals a unexpected shift, as experiments have shown that code-writing agents can fully automate this process by constructing and synthesizing programs that self-adapt to changing environments.

The experimental method relied on giving the programming agents a task description and the ability to interact directly with a physics simulator, while granting each agent a fixed budget of attempts to build the program before freezing the code and testing it on entirely new cases that did not appear during development.The test compared three advanced configurations: Claude Code with the Opus 5 model, and CodeX with the GPT-5.6 Sol and GPT-6 Astra versions.Across twenty-eight simulation environments from the KinDER and PDDLStream benchmarks, with object densities increased beyond the original standards. A total of 980 generated programs were evaluated on one hundred unprecedented test cases each, amounting to 98 000 complete evaluation rounds.

The results showed a sweeping advantage for programming agents over manually engineered plans and one-click generation, achieving success rates ranging from 56 % to 95 % compared with only 47 % for traditional plans in the sixteen environments where those plans were available. More important than raw success rates is system behavior as complexity increases; while traditional plans broke down as the number of objects in the workspace grew, the programs produced by the agents maintained their efficiency and even reduced the computational load per case by an entire order of magnitude relative to classical solutions.

Operational logs revealed that the source of the superiority does not stem from linguistic magic but from experimental interaction.where the agents used simulation capability to self-calibrate physical models, test edge cases and motion limits, and refine release and guidance strategies step by step before finalizing the code. This pattern means that generative programming is no longer merely a tool for completing lines in text editors; it has become a compositional engine that derives motion rules from digital experimentation, and the researchers have published the experiment code and full guidance logs to document the approach.

This shift has immediate relevance for leaders of automation teams, logistics services and smart distribution warehouses in the Gulf and Egypt. Eliminating manually crafted motion plans and generating control programs through simulation reduces operational computing costs by a factor of ten, making it possible to run robot fleets on cheaper on-premise servers within facilities without needing massive cloud computing infrastructure for each maneuver. Moreover, the locally required engineering skill moves from manual mathematical tuning of motion constraints to constructing accurate simulation environments that allow agents to test and autonomously refine their solutions before deploying them on the warehouse and factory floor.