Language models as zero-shot planners with environment feedback

Project mentor: Maitreya Patel

Under the guidance of esteemed Dr.Chitta Baral.

Objective

Use GPT-3 to decompose high-level tasks to low-level executable instructions for an agent in the VirtualHome environment. Compare the decomposition results between ours and the baseline.

Approach

We divided our approach into two sequential steps. First, we check if GPT-3 can give the correct low-level instructions with a minimal prompt (no information about the environment given). Secondly, if GPT-3 fails to give the correct set of instructions, we supply it with a reduced set of objects and permissible actions along with an example set of instructions in the prompt. Since GPT-3 failed to give executable instructions with minimal prompts, we had to execute the second part of the plan.

Setup

1. Clone virtualhome repo from this link
2. Run pip install -r requirements.txt

SEE do_task.ipynb for detailed overview