MAPF_G2RL is an open-source research framework that bridges graph theory and deep reinforcement learning to tackle the multi-agent path finding (MAPF) problem. It encodes nodes and edges into vector representations, defines spatial and collision-aware reward functions, and supports various RL algorithms such as DQN, PPO, and A2C. The framework automates scenario creation by generating random graphs or importing real-world maps, and orchestrates training loops that optimize policies for multiple agents simultaneously. After learning, agents are evaluated in simulated environments to measure path optimality, makespan, and success rates. Its modular design allows researchers to extend core components, integrate new MARL techniques, and benchmark against classical solvers.

Mava is a JAX-based open-source library for developing, training, and evaluating multi-agent reinforcement learning systems. It offers pre-built implementations of cooperative and competitive algorithms such as MAPPO and MADDPG, along with configurable training loops that support single-node and distributed workflows. Researchers can import environments from PettingZoo or define custom environments, then use Mava’s modular components for policy optimization, replay buffer management, and metric logging. The framework’s flexible architecture allows seamless integration of new algorithms, custom observation spaces, and reward structures. By leveraging JAX’s auto-vectorization and hardware acceleration capabilities, Mava ensures efficient large-scale experiments and reproducible benchmarking across various multi-agent scenarios.

MAGAIL implements a multi-agent extension of Generative Adversarial Imitation Learning, enabling groups of agents to learn coordinated behaviors from expert demonstrations. Built in Python with support for PyTorch (or TensorFlow variants), MAGAIL consists of policy (generator) and discriminator modules that are trained in an adversarial loop. Agents generate trajectories in environments like OpenAI Multi-Agent Particle Environment or PettingZoo, which the discriminator uses to evaluate authenticity against expert data. Through iterative updates, policy networks converge to expert-like strategies without explicit reward functions. MAGAIL’s modular design allows customization of network architectures, expert data ingestion, environment integration, and training hyperparameters. Additionally, built-in logging and TensorBoard visualization facilitate monitoring and analysis of multi-agent learning progress and performance benchmarks.

jason-RL adds a reinforcement learning layer to the Jason multi-agent framework, allowing AgentSpeak BDI agents to learn action-selection policies via reward feedback. It implements Q-learning and SARSA algorithms, supports configuration of learning parameters (learning rate, discount factor, exploration strategy), and logs training metrics. By defining reward functions in agent plans and running simulations, developers can observe agents improve decision making over time, adapting to changing environments without manual policy coding.