Plan with Language, Act with Parameter

This repository is based on microRTS and microRTS-Py v0.6.1.

Plan with Language, Act with Parameter (PLAP) is a novel framework that uses predefined parameterized skills as an intermediary level to bypass the challenges of coordinating different levels in complex sequential decision-making environments. The predefined skills can be created by non-expert users, without requiring extensive experience or knowledge. Essentially, we similarly conceal the lowlevel action space and leverage the advantage of the high-level abstract language space, but we additionally expose the parameter domain to LLM agents.

🕹️ Get Started

Prerequisites:

• Python 3.9+
• Poetry
• Java 8.0+
• FFmpeg (for video recording utilities)

# The `poetry install` command above creates a virtual environment for us, in which all the dependencies are installed.
poetry install

# We can use `poetry shell` to create a new shell in which this environment is activated.
poetry shell

🤖 Ai bot vs AI bot

The following command will run a rule-based AI vs rule-based AI example:

cd PLAP

python aibot_vs_aibot.py --blue coacAI --red workerRushAI

Adding --record_video to the python command will record the game process as a video and save it. However, the prerequisite is to set the environment variables.

export DISPLAY=YOUR_IP_ADDRESS:PORT

python aibot_vs_aibot.py --blue coacAI --red workerRushAI --record_video

🎮 LLM vs Ai bot

The following command will run an example of LLM vs. rule-based AI with video recording enabled:

export OPENAI_API_KEY=YOUR_API_KEY

python llm_vs_aibot.py --blue gpt-4o --blue_prompt zero_shot --red randomBiasedAI --record_video

Our few shot with tips PLAP can beat COAC, the winner of the 2020 competition.

qwen_vs_coac_speedup.mp4

⚡ LLM vs LLM

Of course we can also start LLM vs LLM battle：

python llm_vs_llm.py --blue gpt-4o --red gpt-4o-mini --record_video

📁 Directory Structure Description

- gym_microrts/  # microRTS environment
- misc/
- PLAP/  # our project
  - configs/  # experiment configuration
  - grounding/  # grounding LLM to the original environment
    - __init__.py
    - obs2text.py  # convert observation tensor to structured text
    - script_mapping.py  # mapping LLM outputs to action tensors
    - task2actions.py  # parameterized skills
  - llm_agents/  # wrapping LLM
  - utils/
  - aibot_vs_aibot.py  # rule-based AI bot vs rule-based AI bot
  - llm_vs_aibot.py  # LLM vs rule-based AI bot
  - llm_vs_llm.py  # LLM vs LLM
- build.sh  # Compiling Java microRTS
- poetry.lock
- pyproject.toml
- README.md