The MARL Simulator is designed to facilitate efficient and scalable development of multi-agent reinforcement learning (MARL) algorithms. Leveraging PyTorch's distributed backend, it allows users to run parallel training across multiple GPUs or nodes, significantly reducing experiment runtime. The simulator offers a modular environment interface that supports standard benchmark scenarios—such as cooperative navigation, predator-prey, and grid world—as well as user-defined custom environments. Agents can utilize various communication protocols to coordinate actions, share observations, and synchronize rewards. Configurable reward and observation spaces enable fine-grained control over training dynamics, while built-in logging and visualization tools provide real-time insights into performance metrics.

This repository implements emergent communication in multi-agent reinforcement learning using PyTorch. Users can configure sender and receiver neural networks to play referential games or cooperative navigation, encouraging agents to develop a discrete or continuous communication channel. It offers scripts for training, evaluation, and visualization of learned protocols, along with utilities for environment creation, message encoding, and decoding. Researchers can extend it with custom tasks, modify network architectures, and analyze protocol efficiency, fostering rapid experimentation in emergent agent communication.

Vanilla Agents is a lightweight PyTorch-based framework that delivers modular and extensible implementations of core reinforcement learning agents. It supports algorithms like DQN, Double DQN, PPO, and A2C, with pluggable environment wrappers compatible with OpenAI Gym. Users can configure hyperparameters, log training metrics, save checkpoints, and visualize learning curves. The codebase is organized for clarity, making it ideal for research prototyping, educational use, and benchmarking new ideas in RL.