# ChatGPT for Robotics: Design Principles and Model Abilities

**Authors:** Sai Vemprala, Rogerio Bonatti, Arthur Bucker, Ashish Kapoor  
**arXiv:** [2306.17582](https://arxiv.org/abs/2306.17582)  
**Date:** 2023-02-20  
**Summary page:** https://www.summarizepaper.com/en/paper/2306.17582v1/

## Key points

- Experimental study on the use of OpenAI's ChatGPT for robotics applications
- Proposal of a strategy combining prompt engineering design principles and high-level function library
- Focus on evaluating different prompt engineering techniques and dialog strategies for executing robotics tasks
- Exploration of ChatGPT's abilities in free-form dialog, parsing XML tags, synthesizing code, using task-specific prompting functions, and employing closed-loop reasoning through dialogues
- Evaluations across various tasks within the robotics domain, including logical reasoning, geometrical reasoning, mathematical reasoning, aerial navigation, manipulation, and embodied agents
- Effectiveness of ChatGPT in solving several tasks primarily through natural language instructions
- Introduction of PromptCraft as an open-sourced research tool for effective prompting schemes in robotics applications
- Integration of a sample robotics simulator with ChatGPT to facilitate user engagement

## Summary

This paper presents an experimental study on the use of OpenAI's ChatGPT for robotics applications. The authors propose a strategy that combines prompt engineering design principles and the creation of a high-level function library to enable ChatGPT to adapt to various robotics tasks, simulators, and form factors. The study focuses on evaluating different prompt engineering techniques and dialog strategies for executing different types of robotics tasks. The authors explore ChatGPT's ability to engage in free-form dialog, parse XML tags, synthesize code, use task-specific prompting functions, and employ closed-loop reasoning through dialogues. They conduct evaluations across a range of tasks within the robotics domain, including logical reasoning, geometrical reasoning, mathematical reasoning, aerial navigation, manipulation, and embodied agents. The results demonstrate that ChatGPT is effective in solving several of these tasks while primarily interacting with users through natural language instructions. Additionally, the authors introduce PromptCraft – an open-sourced research tool – which provides a platform for researchers to collaboratively upload and vote on examples of effective prompting schemes for robotics applications. PromptCraft also includes a sample robotics simulator integrated with ChatGPT to facilitate user engagement with ChatGPT for robotics applications.