# ProgPrompt: Generating Situated Robot Task Plans using Large Language Models

**Authors:** Ishika Singh, Valts Blukis, Arsalan Mousavian, Ankit Goyal, Danfei Xu, Jonathan Tremblay, Dieter Fox, Jesse Thomason, Animesh Garg  
**arXiv:** [2209.11302](https://arxiv.org/abs/2209.11302)  
**Date:** 2022-09-22  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2209.11302v1/

## Key points

- Authors propose using large language models (LLMs) for task planning in robots
- Existing methods have limitations such as enumerating all possible next steps or generating free-form text
- Authors introduce a programmatic LLM prompt structure to overcome these limitations
- Program-like specifications and example programs are incorporated into the prompt structure for improved performance
- Ablation experiments conducted to determine optimal prompt structure and generation constraints
- State-of-the-art success rates achieved in VirtualHome household tasks
- Method deployed on a physical robot arm for tabletop tasks
- More information available on progprompt.github.io website

## Summary

In their paper titled "ProgPrompt: Generating Situated Robot Task Plans using Large Language Models," authors Ishika Singh, Valts Blukis, Arsalan Mousavian, Ankit Goyal, Danfei Xu, Jonathan Tremblay, Dieter Fox, Jesse Thomason and Animesh Garg address the challenge of task planning for robots. To simplify this process they propose using large language models (LLMs) to score potential actions and generate action sequences based on natural language instructions. Existing methods either require enumerating all possible next steps for scoring or generate free-form text that may include actions not feasible for a given robot in its current context. To overcome these limitations the authors introduce a programmatic LLM prompt structure that enables plan generation across different environments, robot capabilities and tasks. This approach incorporates program-like specifications and example programs into the prompt structure to achieve improved performance across various environments and tasks. The authors conduct ablation experiments to determine optimal prompt structure and generation constraints. They demonstrate state-of-the-art success rates in VirtualHome household tasks and deploy their method on a physical robot arm for tabletop tasks. More information about their work can be found on their website at progprompt.github.io. Overall, this research presents a novel approach to generating situated robot task plans using large language models.