This project provides official code for "Language Models as Zero-Shot Planners," enabling Large Language Models (LLMs) like GPT-3 and Codex to generate action plans for complex tasks without fine-tuning. It targets researchers and developers working with embodied agents or AI planning, offering a method to extract actionable knowledge for task execution.

How It Works

The approach leverages LLMs as zero-shot planners by prompting them with task descriptions and a set of available actions. The LLM generates a sequence of actions to achieve the goal. This method is advantageous as it requires no task-specific training data, relying solely on the LLM's pre-existing knowledge and planning capabilities.

Quick Start & Requirements

• Install: Clone the repository, create a conda environment (conda create --name language-planner-env python=3.6.13), activate it, and install dependencies (pip install -r requirements.txt).
• Prerequisites: Python 3.6.13, CUDA 11.3.
• Usage: Refer to demo.ipynb for a walkthrough.
• Notes: Larger LLMs (OpenAI API, Huggingface Transformers) yield better results. Tuning sampling hyperparameters may be necessary. The code uses actions defined in available_actions.json and examples from available_examples.json.

Highlighted Details

• Demonstrates zero-shot planning capabilities of LLMs for complex tasks.
• Supports OpenAI API and Huggingface Transformers models.
• Utilizes a predefined set of actions for embodied agents, extensible to new domains.
• Provides example prompts and action lists for household tasks.

Maintenance & Community

The project is associated with authors from UC Berkeley, Carnegie Mellon University, and Google Brain. No specific community channels or roadmap are mentioned in the README.

Licensing & Compatibility

The README does not explicitly state a license. The code is provided for research purposes. Compatibility with commercial or closed-source applications is not specified.

Limitations & Caveats

The project requires specific Python (3.6.13) and CUDA (11.3) versions, which may be outdated. Performance is highly dependent on the chosen LLM and careful tuning of sampling hyperparameters. The action set is currently limited to what's defined in available_actions.json, requiring manual updates for new task domains.