# LLM A*: Human in the Loop Large Language Models Enabled A* Search for Robotics

**Authors:** Hengjia Xiao, Peng Wang  
**arXiv:** [2312.01797](https://arxiv.org/abs/2312.01797)  
**Date:** 2023-12-04  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2312.01797v2/

## Key points

- Authors Hengjia Xiao and Peng Wang explore the use of Large Language Models (LLMs) in path planning for mobile embodied agents such as robots.
- The study introduces a novel framework called LLM A*, which combines LLMs with the A* algorithm for efficient and near-optimal path planning with minimal data input.
- Prompts are used to provide essential information to LLMs and enable human feedback on intermediate planning outcomes, making LLM A* transparent and understandable to humans.
- LLM A* enables code-free path planning, promoting accessibility and inclusivity in AI applications, particularly benefiting communities with limited coding proficiency.
- Comparative analysis shows that LLM A* is more efficient in search space utilization compared to traditional methods like A* and Reinforcement Learning (RL), while achieving comparable path outcomes to A* and surpassing RL performance.
- The interactive nature of LLM A* makes it a promising tool for collaborative human-robot tasks, enhancing its practical applicability.
- This research provides insights into leveraging LLMs for enhanced path planning and highlights the potential of integrating human feedback within AI systems for improved decision-making processes.

## Summary

In their research titled "LLM A*: Human in the Loop Large Language Models Enabled A* Search for Robotics," authors Hengjia Xiao and Peng Wang explore the use of Large Language Models (LLMs) in path planning for mobile embodied agents such as robots. The study introduces a novel framework called LLM A*, which combines the commonsense capabilities of LLMs with the utility-optimal A* algorithm to facilitate efficient and near-optimal path planning with minimal data input. Central to this approach is the use of prompts, which serve two purposes: providing essential information to LLMs and enabling human feedback on intermediate planning outcomes. This emphasis on transparency throughout the planning process makes LLM A* akin to a 'white box' system that allows humans to understand and engage with the decision-making mechanisms employed by LLMs. Additionally, by enabling code-free path planning, this approach promotes accessibility and inclusivity in artificial intelligence applications, particularly benefiting communities with limited coding proficiency. Comparative analysis against traditional methods like A* and Reinforcement Learning (RL) showcases the efficiency of LLM A* in terms of search space utilization while achieving comparable path outcomes to A* and surpassing RL performance. The interactive nature of LLM A* also positions it as a promising tool for collaborative human-robot tasks, further enhancing its practical applicability. Overall, this research not only provides valuable insights into leveraging LLMs for enhanced path planning but also highlights the potential for integrating human feedback within AI systems for improved decision-making processes. For those interested in exploring further, codes and supplemental materials related to LLM A* can be accessed on GitHub at https://github.com/speedhawk/LLM-A-.