Robot Motion Planning With Large Language Models:language-to-Paths Guidance Over Superfacets

Robots operating in cluttered environments must compute collision-free motions under nonholonomic dynamics while maintaining fast, reliable planning. This article presents a language-guided motion-planning framework built on a visibility-aware superfacet decomposition, which partitions free space into navigable regions connected by a region graph. A large language model (LLM) converts natural-language preferences into a constrained domain-specific language (DSL) containing route, waypoint, constraint, and parameter hints. Each hint is validated by an affordance/feasibility gate that enforces connectivity and visibility, rejects unreachable suggestions, accounts for recent failures, and bounds updates. Validated hints softly bias inter-region costs and sampling priorities, while kinodynamic rollouts and collision checking determine final feasibility. Experiments with car-like and snake-like robots in cluttered 2D scenes show an approximately 18% reduction in mean planning time and fewer dead-end explorations relative to the superfacet baseline. The framework maintains comparable success rates and path quality, demonstrating that natural-language guidance can be integrated without increasing computational overhead. This makes the framework practical for non-expert operators to direct robot navigation without knowledge of the underlying planning mechanics.