LLM-Advisor: An LLM Benchmark for Cost-efficient Path Planning across Multiple Terrains
Ling Xiao, Toshihiko Yamasaki
TL;DR
The paper addresses cost-efficient path planning across diverse terrains for autonomous robots, where terrain-dependent costs must be minimized beyond simple distance. It introduces LLM-Advisor, a prompt-based system that acts as an advisor to post-process paths produced by A*, RRT*, or LLM-based planners, using terrain-aware prompts and descriptive path representations to guide improvements while mitigating hallucinations. Experimental results on the MultiTerraPath and RUGD_v2 datasets show that LLM-Advisor improves a substantial fraction of paths, with relative improvement (RP) of 70.59% for A*, 69.47% for RRT*, and 78.70% for LLM-A*. The work also demonstrates the superiority of the advisory approach over directly integrating LLMs into the planning loop, and discusses limitations of zero-shot GPT-4o planning and the need for mitigation strategies. Overall, LLM-Advisor offers a practical, scalable way to incorporate global, model-based insights into local planners for energy-aware navigation.
Abstract
Multi-terrain cost-efficient path planning is a crucial task in robot navigation, requiring the identification of a path from the start to the goal that not only avoids obstacles but also minimizes travel costs. This is especially crucial for real-world applications where robots need to navigate diverse terrains in outdoor environments, where recharging or refueling is difficult. However, there is very limited research on this topic. In this paper, we develop a prompt-based approach, LLM-Advisor, which leverages large language models (LLMs) as effective advisors for path planning. The LLM-Advisor selectively provides suggestions, demonstrating its ability to recognize when no modifications are necessary. When suggestions are made, 70.59% of the paths suggested for the A* algorithm, 69.47% for the RRT* algorithm, and 78.70% for the LLM-A* algorithm achieve greater cost efficiency. Since LLM-Advisor may occasionally lack common sense in their suggestions, we propose two hallucination-mitigation strategies. Furthermore, we experimentally verified that GPT-4o performs poorly in zero-shot path planning, even when terrain descriptions are clearly provided, demonstrating its low spatial awareness. We also experimentally demonstrate that using an LLM as an advisor is more effective than directly integrating it into the path-planning loop. Since LLMs may generate hallucinations, using LLMs in the loop of a search-based method (such as A*) may lead to a higher number of failed paths, demonstrating that our proposed LLM-Advisor is a better choice.