Farama-Foundation/MPE2

A set of communication oriented environments

What it solves

MPE2 provides a set of communication-oriented simulation environments for training and testing multi-agent reinforcement learning (MARL) algorithms. It addresses the need for standardized, lightweight environments where agents must learn to interact, communicate, and coordinate to solve tasks in both cooperative and adversarial settings.

How it works

The project implements a series of "particle" environments where agents are represented as circles that can move, see other agents, and interact with static landmarks. It supports both discrete (default) and continuous action spaces, allowing agents to control their movement in cardinal directions or broadcast messages over communication channels. The system tracks observations (position, velocity, and received messages) and rewards based on the agents' success in achieving goals, such as pushing objects or tagging other agents.

Who it’s for

It is designed for researchers and developers working on multi-agent AI, specifically those focusing on emergent communication, coordination, and competitive/cooperative game theory in reinforcement learning.

Highlights

  • Diverse Interaction Models: Includes both cooperative environments (e.g., Simple Spread, Simple Formation) and adversarial ones (e.g., Simple Tag, Simple Push).
  • Flexible Action Spaces: Supports both discrete and continuous control for movement and communication.
  • Communication Channels: Built-in mechanisms for agents to broadcast and receive messages to coordinate actions.
  • Improved Stability: Based on OpenAI's original MPE, but with fixes for reward consistency and discrete action space defaults.

Related

  • Project
    AgentR1/Agent-R1

    A modular framework for Agentic Reinforcement Learning that trains multi-step LLM agents by treating each interaction turn as a step-level MDP transition.

  • Project
    Farama-Foundation/MAgent2

    A library for creating pixel-based gridworld environments where large numbers of agents interact in competitive scenarios to study collective intelligence.

  • Project
    TsinghuaC3I/MARTI

    An open-source framework for training LLM-based multi-agent systems using reinforcement learning, featuring tree search augmentation for complex reasoning and code generation.

  • Dispatch
    Neural MMO: A Massively Multiagent Game Environment

    OpenAI has released Neural MMO, a persistent, large-scale environment for reinforcement learning agents that demonstrates how increased population size and species diversity drive exploration and niche formation.

  • Project
    facebookresearch/Pearl

    A production-ready Reinforcement Learning library from Meta for building modular AI agents capable of handling complex, real-world environments with sparse feedback and dynamic action spaces.