TorchRL + PettingZoo MPE in 60 Minutes
TL;DR: Build a reproducible multi-agent reinforcement learning (MARL) pipeline using TorchRL and PettingZoo's multi-agent particle environment (MPE), train using centralized training, decentralized execution (CTDE), and verify that communication is meaningful using structured diagnostics and causal ablations.
Tutorial in 30 Seconds#
TorchRL + PettingZoo MPE is a powerful stack for building and verifying multi-agent reinforcement learning systems with measurable coordination.
Key capabilities:
- Multi-agent environment integration: Clean API for wiring PettingZoo environments with TorchRL's PyTorch-native RL pipeline
- Centralized training, decentralized execution (CTDE): Stable training with decentralized per-agent actors and a centralized critic
- Communication verification: Structured diagnostics including message entropy, message change rate, and observation-derived verification checks
- Causal ablations: Built-in support for testing communication importance through full communication, disabled communication, and random communication modes
- Outcome-aligned evaluation: Binary success rates, goal-distance debugging, and structured communication metrics
This tutorial's goal is to show you in 60 minutes:
- The core architecture for building MARL systems with TorchRL and PettingZoo
- How to set up CTDE training with centralized critics and decentralized actors
- How to verify coordination and communication meaning through structured diagnostics and causal ablations
- Why training curves alone are not sufficient proof of cooperation
The Problem with Most MARL Tutorials#
Most multi-agent reinforcement learning tutorials show:
- Moving loss curves
- Increasing rewards
- "Healthy-looking" training
But they rarely verify:
- Whether agents truly coordinate
- Whether communication carries useful information
- Whether evaluation-time success improves
In multi-agent RL, training curves are not proof of cooperation. This tutorial focuses on measurable coordination.
Official References#
- TorchRL: A PyTorch Reinforcement Learning Library
- PettingZoo: Parallel Multi-Agent Environment Library
- Multi-Agent Particle Environment (MPE)
Tutorial Content#
This tutorial includes all the code, notebooks, and Docker containers in tutorials/TorchRL_Multi_Agent_Cooperation
README.md: Instructions and setup for the tutorial environment- A Docker system to build and run the environment using our standardized approach
- Environment and wrapper components:
- PettingZoo MPE task wrappers (e.g.,
simple_reference) - TorchRL-compatible wrappers
- Explicit message-channel wiring
- CTDE Training setup:
- Decentralized per-agent actors
- Centralized critic
- Stable policy optimization
- Evaluation and diagnostics:
- Binary success rate metrics
- Goal-distance debugging
- Structured communication metrics including
message_entropyandmessage_change_rate - Causal ablations:
full_comm: Standard communication enableddisable_comm: Communication disabledrandom_comm: Random communication