# LanguageMPC: Large Language Models as Decision Makers for Autonomous Driving

**Authors:** Hao Sha, Yao Mu, Yuxuan Jiang, Li Chen, Chenfeng Xu, Ping Luo, Shengbo Eben Li, Masayoshi Tomizuka, Wei Zhan, Mingyu Ding  
**arXiv:** [2310.03026](https://arxiv.org/abs/2310.03026)  
**Date:** 2023-10-04  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2310.03026v1/

## Key points

- Existing learning-based autonomous driving (AD) systems face challenges in understanding high-level information, generalizing to rare events, and providing interpretability.
- Large Language Models (LLMs) are used as decision-making components for complex AD scenarios that require human commonsense understanding.
- Cognitive pathways are developed to enable comprehensive reasoning with LLMs.
- Algorithms are devised for translating LLM decisions into actionable driving commands.
- Extensive experiments show that the proposed method consistently outperforms baseline approaches in single-vehicle tasks and effectively handles complex driving behaviors including multi-vehicle coordination.
- The success is attributed to the commonsense reasoning capabilities of LLMs.
- This research represents an initial step towards leveraging LLMs as effective decision-makers for intricate AD scenarios in terms of safety, efficiency, generalizability, and interoperability.

## Summary

, , , , 

Existing learning-based autonomous driving (AD) systems face challenges in understanding high-level information, generalizing to rare events, and providing interpretability. To address these issues, this research utilizes Large Language Models (LLMs) as decision-making components for complex AD scenarios that require human commonsense understanding. The authors develop cognitive pathways to enable comprehensive reasoning with LLMs and devise algorithms for translating LLM decisions into actionable driving commands. Through this approach, LLM decisions seamlessly integrate with low-level controllers through guided parameter matrix adaptation. Extensive experiments demonstrate that the proposed method consistently outperforms baseline approaches in single-vehicle tasks and effectively handles complex driving behaviors including multi-vehicle coordination. This success is attributed to the commonsense reasoning capabilities of LLMs. The paper represents an initial step towards leveraging LLMs as effective decision-makers for intricate AD scenarios in terms of safety, efficiency, generalizability, and interoperability. The authors hope their work will inspire future research in this field.