[TOC]

  1. Title: ProgPrompt: program generation for situated robot task planning using large language models
  2. Author: Ishika Singh et. al.
  3. Publish Year: 28 August 2023
  4. Review Date: Mon, Jan 29, 2024
  5. url: https://progprompt.github.io/

Summary of paper

Motivation

Classical Task planning

  • requires myriad domain knowledge
  • large serach space, hard toscale
  • domain specific
  • require concrete goal specification

Planning with LLMs

image-20240129205334711

  • LLM is not situated in the scene
  • Plan steps using unavailable actions and objects
  • Text-to-robot action mapping may not be trivial
  • combinatorial admissible action space.

Contribution

  • present a programmatic LLM prompt structure that enables plan generation function across situated environments, robot capabilities and tasks

image-20240129205546614