Can Large Language Models Solve Robot Routing?

TL;DR

This work probes whether large language models can fully replace the routing optimization pipeline by translating natural language tasks into robot routes. It constructs a dataset of 80 problems across 8 routing variants and evaluates three LLM-based frameworks—single attempt, self-debugging, and self-debugging with self-verification—across four LLMs and varying contextual information. Key findings show that self-debugging and self-verification increase solution feasibility but do not consistently reduce the optimality gap, with context effects being highly task- and model-dependent. The study identifies practical challenges, notably verifier reliability and context sensitivity, and suggests directions for building on existing LLMs or fine-tuning them to improve end-to-end performance in robot routing.

Abstract

Routing problems are common in mobile robotics, encompassing tasks such as inspection, surveillance, and coverage. Depending on the objective and constraints, these problems often reduce to variants of the Traveling Salesman Problem (TSP), with solutions traditionally derived by translating high-level objectives into an optimization formulation and using modern solvers to arrive at a solution. Here, we explore the potential of Large Language Models (LLMs) to replace the entire pipeline from tasks described in natural language to the generation of robot routes. We systematically investigate the performance of LLMs in robot routing by constructing a dataset with 80 unique robot routing problems across 8 variants in both single and multi-robot settings. We evaluate LLMs through three frameworks: single attempt, self-debugging, and self-debugging with self-verification and various contexts, including mathematical formulations, pseudo-code, and related research papers. Our findings reveal that both self-debugging and self-verification enhance success rates without significantly lowering the optimality gap. We observe context-sensitive behavior - providing mathematical formulations as context decreases the optimality gap but significantly decreases success rates and providing pseudo-code and related research papers as context does not consistently improve success rates or decrease the optimality gap. We identify key challenges and propose future directions to enhance LLM performance in solving robot routing problems. Our source code is available on the project website: https://sites.google.com/view/words-to-routes/.

Figures (9)

• Figure 1: Two illustrative examples of robot routing problems. (left) a marine robot needs to visit a set of candidate locations to take ocean samples, and the operator wants the robot to finish the sampling process as fast as possible - this corresponds to a Traveling Salesman Problem (TSP); (right) an aerial robot with cameras needs to visit a subset of all the candidate viewpoints to ensure sufficient coverage of the bridge, which corresponds to the generalized TSP.
• Figure 2: An illustrative scenario based on the Capacitated Vehicle Routing Problem (CVRP). Left: an overview of acquiring solutions given task descriptions in a zero-shot manner. Right: solution visualization.
• Figure 3: Framework overview. (a) single attempt; (b) self-debugging; (c) self-debugging with self-verification. In (c), the solution generator and the verifier generator have the same structure as that in (b).
• Figure 4: Overview of the prompt template.
• Figure 5: CVRP result from self-debugging only framework and that from self-debugging with self-verification framework for P-n21-k2. The results from the self-debugging framework provide an infeasible solution that misses location 2, while by adding self-verification, the framework identifies the incorrect solution and fixes it.