# Translating Natural Language to Planning Goals with Large-Language Models

**Authors:** Yaqi Xie, Chen Yu, Tongyao Zhu, Jinbin Bai, Ze Gong, Harold Soh  
**arXiv:** [2302.05128](https://arxiv.org/abs/2302.05128)  
**Date:** 2023-02-10  
**Summary page:** https://www.summarizepaper.com/en/paper/2302.05128v1/

## Key points

- Large language models (LLMs) have potential applications in various domains of natural language processing (NLP)
- LLMs struggle with accurate reasoning and planning, limiting their usefulness for robotics-related tasks
- The study investigates whether LLMs can translate goals expressed in natural language into a structured planning language
- Experiments using GPT 3.5 variants reveal that LLMs are better suited for translation than planning
- LLMs can leverage commonsense knowledge and reasoning to fill in missing details from under-specified goals commonly found in natural language
- However, LLMs struggle with generating goals involving numerical or physical reasoning and are sensitive to the prompts used
- Caution should be exercised when using LLMs for real-world applications such as robotics
- Developing more robust NLP models capable of accurate reasoning and planning is important for real-world applications

## Summary

The recent development of large language models (LLMs) has sparked excitement about their potential applications across various domains in natural language processing (NLP). However, research has shown that LLMs struggle with accurate reasoning and planning, which may limit their usefulness for robotics-related tasks. This study investigates whether LLMs can translate goals expressed in natural language into a structured planning language. Experiments using GPT 3.5 variants reveal that LLMs are better suited for translation than planning. They found that LLMs can leverage commonsense knowledge and reasoning to fill in missing details from under-specified goals commonly found in natural language. However, they also discovered that LLMs struggle with generating goals involving numerical or physical reasoning and are sensitive to the prompts used. Thus, while these models show promise for translating natural language to structured planning languages, caution should be exercised when using them. This study highlights the importance of developing more robust NLP models capable of accurate reasoning and planning for real-world applications such as robotics.