LLM Google FootballMoonshot Guide

Let Large Language Models take over the decision-making brain on the football pitch, enabling pure text-based end-to-end tactical execution and bulk parallel evaluation.

Quick Start

What is LLM-Football?

🧠

👁 Situational Awareness Translation (Obs-to-Text)

No need to process complex pixel streams. Converts GRF's raw 3D coordinates into aligned natural language, letting the LLM instantly understand the pitch situation.

🎮

⚡ Parallel Leaderboard Testing

Built-in one-click racing script run_multiple_experiments.py. Easily configure 4 mainstream LLMs and automatically evaluate them against each other.

🏆

📊 Cool Visual Reports

Generates and deploys a multi-dimensional tournament Leaderboard highlighting high-value metrics like win rate, average advance score, and latency.

Can Large Language Models Really Play Football?

In traditional Reinforcement Learning (RL), training a game AI often consumes tens of millions of frames of experience fragments. In the LLM era, leveraging world knowledge and Zero-shot reasoning capabilities, can we use a single Prompt to let a general LLM (like GPT-4 / Qwen / GLM / Kimi) directly command players to run, shoot, and win matches?

This project provides an out-of-the-box experimental framework, currently supporting the academy_3_vs_1_with_keeper scenario to automatically evaluate the combat power of various AI contestants.