Researchers have made a significant step forward in helping robots and automated systems plan and execute complex tasks involving both decision-making and physical movement. The new study explores how “coding agents”—artificial intelligence systems that write their own programs—can learn to solve generalized task and motion planning (TAMP) problems more efficiently than traditional hand-crafted methods. This advancement matters because it could enable smarter robots that adapt quickly to new situations without extensive manual programming, potentially impacting fields from manufacturing to autonomous vehicles.
Key Takeaways
- Coding agents were tested on 28 different simulated environments involving tasks requiring both planning discrete steps and managing physical constraints.
- These agents generated programs that successfully handled new, unseen instances with success rates between 56% and 95%, outperforming hand-engineered planners which averaged 47% success.
- As the complexity of tasks increased, especially with more objects involved, coding agents maintained higher success rates while using roughly ten times less computation per task than traditional planners.
- Analysis of the agents’ behavior showed they actively interacted with environments to calibrate physical models and refine their strategies, indicating a level of adaptive problem-solving.
Task and motion planning (TAMP) is a challenging area in robotics and AI because it involves both deciding what discrete actions to take (like picking up an object or opening a door) and ensuring those actions respect physical constraints such as geometry, movement limits, and dynamics. Traditional approaches often require expert-designed algorithms tailored to specific problems, which can be time-consuming and inflexible.
The researchers investigated whether coding agents—AI systems capable of writing computer programs—could automatically synthesize generalized solutions that work across many different but related tasks. Instead of manually engineering each planner, these agents receive a description of the task and access to a simulator that models the environment. Within a fixed “synthesis budget” (a limit on how much time or computational effort they can spend), each agent interacts with the simulator, experimenting and learning how best to solve the problem. The agent then produces a program that can be reused to solve new instances without further modification.
In the experiments, three different coding agents, including models based on advanced large language models like Claude Code (Opus 5) and Codex variants (GPT-5.6 Sol and GPT-6 Astra), were evaluated on a large-scale benchmark consisting of 28 simulated environments derived from KinDER and PDDLStream frameworks. These environments included tasks with varying numbers of objects, increasing the difficulty. Each generated program was tested on 100 new instances, resulting in a total of 98,000 evaluation episodes.
The results showed that coding agents consistently outperformed traditional hand-engineered planners and other baseline methods, including one-shot program generation and an LLM-based generalized planning approach. Notably, the agents’ programs retained strong performance even as tasks became more complex, demonstrating both efficiency and scalability. Logs from the experiments revealed that the agents didn’t just passively generate code—they actively used their interactions with the environment to improve their internal understanding of physical models, test edge cases, and fine-tune their strategies.
These findings suggest that coding agents could become a powerful tool for developing flexible, general-purpose planning systems that require less manual engineering. This could accelerate progress in robotic applications where adaptability and efficiency are crucial, such as warehouse automation, service robots, or autonomous vehicles navigating dynamic environments. The researchers have made their code and prompts publicly available, encouraging further exploration and development in this promising area.
Looking ahead, future work may explore extending these methods to even more complex real-world scenarios, integrating learning with physical robots, or combining coding agents with other AI techniques to further improve robustness and generalization. While challenges remain, this research marks an important milestone in automating the creation of intelligent planning solutions.
Based on research published on arXiv by Matteo Merler, Bowen Li, Josh Roy et al..