Neuro-Symbolic Program Synthesis for Multi-Hop Natural Language Navigation

Solving navigation problems from natural language descriptions is essential for advancing humanrobot interaction and enhancing the usability of autonomous systems. Symbolic approaches to path planning excel in well defined environments but cannot cope with the ambiguity of natural language inputs. On the other hand, neural solutions centered on large language models (LLMs) can parse free-form natural language but lack the reasoning capabilities for solving complex multihop path planning problems. In this paper, we propose a neuro-symbolic framework based on program synthesis for multi-hop natural language navigation called NSPS. The framework uses an LLM to parse the problem definition in natural language, a graph of the environment, and an API of a graph library. Next, the code generation capabilities of the LLM are used to synthesize a program for path planning and verification. The path planning program is executed to generate a solution path that is checked by the verification program. A selfcorrection loop is used to fix both syntax and value errors. The framework is evaluated using 600 multihop navigation tasks with 1 to 10 hops. Compared with neural approaches, the NSPS framework improves the success rate and path efficiency by an average of 64.3% and 19.4% across all tasks, respectively.