LLM A*

LLM A* is a novel algorithm that integrates traditional heuristic algorithms together with large language model (LLM), in order to achieving a lower search complexity and higher quality of planned path in comparison with data-drived path planning algorithms. Based on GPT-3.5-turbo interaction sessions, LLM A* is able to meet a near-A* degree in the light of the matrics of search complexity, path steps and maximum deviation times (MDT) under special obstacle grid-map environments with different sizes.

arXiv | supplement

Requirements & Running

We saved each codenote of LLM A*, LLM Greedy and PPO on GoogleColab as a corresponding gist that can be immediately run. However, an available OpenAI API key is required in LLM-based algorithms. (fix with the code on google colab) please change the code below with your API key before running:

self.openai_key = 'YOUR_KEY'

Moreover, please be aware about the path which should be self-defined in your Google drive. For LLM-based algorithms, you should be care about: The path saving environments For PPO, you should be extraly care about:

1. path of statistic data (about the score and steps)
2. path of network checkpoints

Especially, you can manually modify the number in 'episodes.txt' to 0 when starting a new training.

Potencial Problems

There are possibly two problems during running. When the problems below appear, please just re-run the code:

1. For both LLM A* and LLM Greedy, an error with 'empty arg' is thrown out.
2. For PPO, an error with [nan, nan] is thrown out.

Additionally, considering the fact that there is a certain degree of randomness in the result of LLM A*, LLM Greedy and PPO, it is normal for the reimplementation results to be different from those in the paper.

BiblioTeX

  @misc{xiao2023llm,
        title={LLM A*: Human in the Loop Large Language Models Enabled A* Search for Robotics}, 
        author={Hengjia Xiao and Peng Wang},
        year={2023},
        eprint={2312.01797},
        archivePrefix={arXiv},
        primaryClass={cs.RO}
  }