limitless-ai 0.1.0

Universal AI Game Agent

Limitless: Universal AI Game Agent

Limitless is a high-performance, vision-only foundation model agent designed to learn and play any commercial video game. It uses a ResNet50 backbone to map raw gameplay pixels directly to virtual controller actions.

Features

• Universal Simulator: A Gymnasium-compatible environment that captures any game window and sends inputs via a Virtual Xbox 360 Controller.
• Vision-Only Generalist: No game-specific IDs or embeddings. The model learns pure visual-to-action mapping, allowing for zero-shot potential across different games.
• Robust Extractor: Automatically extracts controller states from gameplay videos, even with semi-transparent overlays (e.g., Cuphead, Ready or Not).

🛠️ Installation

1. Prerequisites

• Python 3.10+
• ViGEmBus Driver: Required for virtual controller support. Download here.
• NVIDIA GPU: Recommended for high-speed training and inference.

2. Setup

# Clone the repository
git clone https://github.com/haumlab/limitless
cd limitless

# Install the package
pip install .

Usage

Once installed, you can use the limitless command directly.

1. Extract Data

Extract actions from your own gameplay videos.

limitless extract --input "path/to/videos" --output "dataset/my_game"

2. Train the Model

Train the "Limitless" model on one or multiple datasets.

limitless train --dataset dataset/readyornot --dataset dataset/cuphead

3. Run the Agent

Run the trained agent on a live game window.

limitless run --window "Cuphead" --model limitless_latest.pth

4. Test on Video

Visualize the model's predictions on a video file before running it live.

limitless test-video --video "gameplay.mp4" --model limitless_latest.pth

Technical Architecture

• Backbone: ResNet50 (Pre-trained on ImageNet).
• Policy Heads:
  • Buttons: Multi-label classification (Sigmoid + BCEWithLogitsLoss).
  • Sticks: Regression (MSELoss) for LX, LY, RX, RY axes.
• Input: 224x224 RGB frames.
• Output: 14 Buttons, 4 Stick Axes.

Safety & Isolation

Limitless is designed to be non-intrusive:

• It uses a Virtual Gamepad (vgamepad) and does not touch your physical mouse or keyboard.
• It captures screen data via mss and does not inject code into game processes.

License

MIT