Meta MMO: A Massive Update to Neural MMO

I’m excited to share a massive update to Neural MMO that integrates all the previous NeurIPS competition know-how and makes it 3x faster! By Joseph Suárez, Ryan Sullivan, and me!

Introducing Meta MMO

A new collection of many-agent minigames for training generalist RL agents.

Highlights

• Diverse minigames with free-for-all & team settings, domain randomization, and adaptive difficulty
• Up to 3X faster training vs Neural MMO 2, allowing 1B step experiments on a gaming desktop
• Successfully trained a generalist agent to play multiple minigames with one set of weights

Open Source

Meta MMO, along with baselines and code, is open-source under MIT license. We hope it catalyzes more progress in multiagent RL.

• Paper: arxiv.org/abs/2406.05071
• Code: github.com/kywch/meta-mmo

Huge Thanks to PufferAI

Huge shout-out to PufferAI & PufferLib, a new super efficient RL library, for powering this work! This success directly led to my later work simplifying PHC with PufferLib.

Game Replays

You can watch these minigame replays in your browser:

Full Config Minigames

• Survival
• Team Battle
• Multi-task Training

Mini Config Minigames

• Team Battle
• Protect the King
• Race to the Center
• King of the Hill
• Sandwich

Context

This work builds on the 2023 NeurIPS Neural MMO Challenge and my talk at NeurIPS 2023.