Hugging Face Snowball Fight ML-Agents Environment

Hugging Face has introduced Snowball Fight 1vs1, the company's first custom Deep Reinforcement Learning (DRL) environment. This release marks the beginning of a broader effort to create an ecosystem for DRL researchers and enthusiasts using the Unity ML-Agents toolkit.

Snowball Fight 1vs1 Environment

Snowball Fight 1vs1 is a game developed using Unity ML-Agents where users can compete by shooting snowballs against a Deep Reinforcement Learning agent. The environment is open source, hosted on the Hugging Face Hub, and available for interactive play via Hugging Face Spaces.

Unity ML-Agents Ecosystem at Hugging Face

Hugging Face is leveraging the Unity Machine Learning Agents Toolkit—an open-source library for building games and simulations to train intelligent agents—to establish a three-part ecosystem:

  • Custom Environment Development: Hugging Face is developing and sharing open-source environments (such as snowball fights, racing, and puzzles) on the Hugging Face Hub to facilitate experimentation with new problems.
  • Hosting and Model Sharing: The platform allows users to host environments, save models, and share them directly on the Hugging Face Hub.
  • Demo Deployment: Users can utilize Hugging Face Spaces to quickly host and showcase their DRL results to the community.

Future Roadmap and Technical Extensions

Hugging Face plans to expand the ML-Agents ecosystem through the following initiatives:

  • Technical Documentation: The release of technical tutorials specifically focused on ML-Agents.
  • Multi-Agent Collaboration: Development of a Snowball Fight 2vs2 version. This version will utilize MA-POCA, a Deep Reinforcement Learning algorithm designed to train cooperative behaviors within teams.
  • Environment Expansion: The creation of additional custom environments to be hosted on the Hugging Face platform.

Sources