When to Plan and When to React: Selective Planning and LLM-Integrated Code Execution in Robot Planning

Abstract

Robots operating in real-world environments must make decisions that are both effective and efficient. Deliberative planning methods reliably find optimal solutions but are computationally expensive, while reactive policies are fast and generalize broadly but may fail in structurally complex regions where search is necessary. This thesis investigates how to combine deliberative planning and reactive policy learning in programmatic policies, where robot behavior is represented as interpretable, executable code. We present an approach in which a large language model (LLM) generates policies that can choose when to use a planner and when to act reactively. We evaluate this approach in two domains using two different planners: a custom discrete maze domain and a continuous Motion2D domain from the KinDER benchmark, paired with a search-based planner and a sampling-based motion planner. In the maze domain, the hybrid approach solved every trial with a 68% reduction in node expansions over pure planning and 31% below a hand-coded oracle. In Motion2D, it achieved 96% success versus 72% for pure planning with 83% fewer collision checks, and was the only approach with non-zero success at the highest difficulty level. Together, these results show that LLM-synthesized programmatic policies can outperform both pure planning and pure reactive execution, and that this framework generalizes across domains, providing a foundation for future work on autonomous decision-making in more complex robotic settings.