Skip to content

Latest commit

 

History

History

examples

README.md

This folder contains the prompts used for task decomposition from natural language in our paper, as well as examples of ChatGPT output. We used gpt-35-turbo (0301) for the experiments. If you are using version 0301, set api_version in aimode.py to '2022-12-01'. For any other version, use '2023-05-15'. By default, api_version is set to '2022-12-01'. If you are using the OpenAI API instead of the Azure OpenAI, set use_azure in aimode.py to False. By default, use_azure is set to True.

We provide five example prompts for the following tasks:

  • task_decomposition: This folder contains the prompts used for task decomposition from natural language.
  • task_decomposition_dual_arm: This folder contains the prompts used for task decomposition from natural language for dual-arm robots.
  • task_decomposition_logic: This folder contains the prompts used for task decomposition with conditional logic from natural language.
  • task_decomposition_virtualhome: This folder contains the prompts used for the experiment with VirtualHome. Please check the README.md in this folder for more details.
  • task_decomposition_virtualhome_supplementary: This folder contains the prompts used for the experiment with VirtualHome (supplementary data). Please check the README.md in this folder for more details.

Directory structure should look like this:

root_folder
│───aimodel.py
├───out/
│───prompt/
│───query/
│───system/
  • aimodel.py: a python script for calling ChatGPT
  • system/: Contains a text file to be inserted at the beginning of the prompt.
  • prompt/: A folder for storing the prompts.
  • query/: Contains a template for converting user input into prompts.
  • out/: A folder for storing the output of ChatGPT.

Please note that:

  • We conducted experiments that involved multi-step natural language instructions in various environments. The environment definitions and natural language instructions are written in aimodel.py.

  • Some of the results shown in out/ include outputs that were re-generated by ChatGPT based on user feedback. Running aimodel.py allows for text feedback via standard input after ChatGPT's response to each natural language prompt.

Microsoft's sample code for using ChatGPT will be helpful for understanding how to use the API.