LLM Dynamic Planner (LLM-DP)

Setup

1. Install alfworld following instructions here.
2. Install requirements: pip install -r requirements.txt.
3. Install docker for running LAPKT planner. There are two different docker images available:
   • The docker image for the linux/arm64 platform is available here. See the Dockerfile for more details.
   • The docker image for the linux/amd64 platform is available here

Config

The config file is a .env file. The following variables are available:

• openai_api_key: OpenAI API key.
• alfworld_data_path: path to the alfworld/data directory.
• alfworld_config_path: path to the alfworld/configs directory.
• pddl_dir: path to the directory where PDDL files are stored.
• pddl_domain_file: name of the PDDL domain file.
• planner_solver: name of the planner solver to use. Currently, only bfs_f and ff are supported.
• planner_timeout: timeout for the planner in seconds (default: 30).
• planner_cpu_count: number of CPUs to use for the planner (default: 4).
• top_n: number of plans to generate (default: 3).
• platform: platform to use for the planner (default: linux/arm64).
• output_dir: path to the output directory (default: output).
• seed: random seed (default: 42).
• name: name of the experiment (default: llmdp).
• llm_model: name of the LLM model to use (default: gpt-3.5-turbo-0613).
• sample: whether to use LLM to instantiate beliefs (default: llm).
• use_react_chat: activate ReAct baseline (default: False).
• random_fallback: activate random fallback (default: False).

Copy the .env.draft to .env and fill the variables in the created .env file with the appropriate values.

See config.py for more details.

Run

Run the following command to run the LLM-DP (or ReAct) agent:

python main.py