DIAMBRA Arena provides a standardized reinforcement learning (RL) platform for researchers and practitioners, offering a suite of arcade fighting game environments with a Python API compatible with OpenAI Gym/Gymnasium. It facilitates experimentation across various RL subfields, including multi-agent, self-play, and imitation learning, by leveraging pixel and RAM state observations.

How It Works

The platform interfaces with popular retro fighting games, providing episodic RL tasks with discrete actions and combined pixel/RAM observations. It supports both single-player and two-player modes, enabling research into competitive multi-agent and human-agent interactions. The design prioritizes compatibility with major RL libraries like Stable Baselines3 and Ray RLlib.

Quick Start & Requirements

• Install: pip install diambra-arena (or diambra-arena[stable-baselines3], diambra-arena[ray-rllib])
• Prerequisites: Docker Desktop, Python 3.x. Requires game ROMs to be downloaded separately and validated.
• Setup: Requires user registration on the diambra.ai website. ROM validation and setup can take time depending on the number of games.
• Links: Documentation, Website

Highlighted Details

• Supports 1P and 2P modes for diverse RL research.
• Natively interfaces with Stable Baselines3 (v2.1.) and Ray RLlib (v2.7.).
• Includes a competition platform with a global leaderboard and agent evaluation.
• Offers examples for single-player, multi-player, wrappers, and imitation learning.

Maintenance & Community

• Active development with community support via Discord.
• Links: Discord, Twitter

Licensing & Compatibility

• Subject to DIAMBRA's Terms of Use.
• Generally compatible with commercial use, but specific terms should be reviewed.

Limitations & Caveats

• Requires users to register on the official website and download/validate game ROMs, which can be a significant hurdle.
• The platform is subject to specific Terms of Use, which users must accept.