Go to ACL ARR 2026 January homepageLLM-Enhanced Path Planning For Large-Scale Grid Maps
Keywords: Path Planning, Large Language Models, Heuristic Search, Large-scale Grid Maps, Navigation, Spatial Reasoning
Abstract: Path planning in grid maps is a fundamental problem in abroad range of applications. Existing works fall into traditional algorithms and learning-based algorithms. The traditional algorithms could generally guarantee completeness or optimality, but demand ultra-high computational and memory cost in large-scale maps. The learning-based ones adopt end-to-end learning paradigms that exploit priors to reduce computation, which however rely on task-specific pretraining and face scalability issues. Recently, Large Language Models (LLMs) have shown strong reasoning capacity for planning tasks, but suffer from spatial illusion and unstable performance in terms of long path length, long runtime and large memory occupation. Observing both the advantages and limitations of the SOTA LLM-enhanced algorithm, we propose iLLM-A*, an innovative LLM-enhanced efficient path planning algorithm for large-scale maps. iLLM-A* addresses the limitations of the SOTA LLM-Enhanced algorithm by integrating $4$ modules: Optimal Standard A*, Waypoint Generating by Incremental Learning-Based LLM, Appropriate Waypoint Selection, and Circle-based Neighboring Waypoint Connection. iLLM-A* leverages the base LLM without fine-tuning, which bears the advantage of scalability. Finally, comprehensive evaluations are conducted to compare iLLM-A* with various baselines, using a set synthetic maps and a set of widely-adopted benchmarking maps related to $6$ scenarios. The results show that iLLM-A* significantly outperforms the baselines.
Paper Type: Long
Research Area: Multimodality and Language Grounding to Vision, Robotics and Beyond
Research Area Keywords: Multimodality and Language Grounding to Vision, Robotics and Beyond
Contribution Types: Approaches low compute settings-efficiency
Languages Studied: English
Submission Number: 738
Loading