This repository provides a platform for training AI agents on a wide array of real-time applications, including games and websites, by interfacing with them through screen pixels and input events. It targets AI researchers and developers seeking to measure and improve general intelligence across diverse, complex environments.

How It Works

Universe operates by packaging applications within Docker containers, presenting a standardized VNC and WebSocket interface to the AI agent. This approach allows interaction with applications without requiring access to their internal code or APIs, treating them as black boxes akin to human users. The core client-side library, built on OpenAI Gym, manages communication with these remote environments.

Quick Start & Requirements

• Install via pip install -e . after cloning the repository.
• Requires Python 2.7 or 3.5, Docker, Go (>= 1.5), numpy, incremental, and libjpeg-turbo.
• Official documentation: https://github.com/openai/universe/blob/master/README.md
• Starter Agent: https://github.com/openai/universe-starter-agent

Highlighted Details

• Offers over 1,000 environments, including Atari, Flash games, and browser-based tasks (World of Bits).
• Supports vectorized environments for controlling multiple instances simultaneously.
• Environments run in real-time, requiring continuous action submission from the agent.
• VNC client is accessible at http://localhost:15900/viewer/?password=openai.

Maintenance & Community

• This repository has been deprecated in favor of openai/retro.
• Community support and further development are directed towards the Retro library.
• Discussion board: https://discuss.openai.com/c/Universe

Licensing & Compatibility

• The repository does not explicitly state a license in the README.
• Compatibility for commercial use or closed-source linking is not specified.

Limitations & Caveats

This project is deprecated, with all development shifted to the openai/retro library. While it supports Linux and OSX, Windows support is experimental. The real-time nature of environments means agents cannot pause, and initial reset() observations may be None.