gym-multigrid provides a suite of customizable gridworld environments designed for multi-room navigation and exploration tasks in reinforcement learning. Each environment consists of interconnected rooms populated with objects, keys, doors, and obstacles. Users can adjust grid size, room configurations, and object placements programmatically. The library supports both full and partial observation modes, offering RGB or matrix state representations. Actions include movement, object interaction, and door manipulation. By integrating it as a Gym environment, researchers can leverage any Gym-compatible agent, seamlessly training and evaluating algorithms on tasks like key-door puzzles, object retrieval, and hierarchical planning. gym-multigrid’s modular design and minimal dependencies make it ideal for benchmarking new AI strategies.

LifelongAgentBench is designed to simulate real-world continuous learning environments, enabling developers to test AI agents across a sequence of evolving tasks. The framework offers a plug-and-play API to define new scenarios, load datasets, and configure memory management policies. Built-in evaluation modules compute metrics like forward transfer, backward transfer, forgetting rate, and cumulative performance. Users can deploy baseline implementations or integrate proprietary agents, facilitating direct comparison under identical settings. Results are exported as standardized reports, featuring interactive plots and tables. The modular architecture supports extensions with custom dataloaders, metrics, and visualization plugins, ensuring researchers and engineers can adapt the platform to varied application domains.

The mario-ai project offers a comprehensive pipeline for developing AI agents to master Super Mario Bros. using neuroevolution. By integrating a Python-based NEAT implementation with the OpenAI Gym SuperMario environment, it allows users to define custom fitness criteria, mutation rates, and network topologies. During training, the framework evaluates generations of neural networks, selects high-performing genomes, and provides real-time visualization of both gameplay and network evolution. Additionally, it supports saving and loading trained models, exporting champion genomes, and generating detailed performance logs. Researchers, educators, and hobbyists can extend the codebase to other game environments, experiment with evolutionary strategies, and benchmark AI learning progress across different levels.

MultiAgentPacman offers a Python-based game environment where users can implement, visualize, and benchmark multiple AI agents in the Pacman domain. It supports adversarial search algorithms like minimax, expectimax, alpha-beta pruning, as well as custom reinforcement learning or heuristic-based agents. The framework includes a simple GUI, command-line controls, and utilities to log game statistics and compare agent performance under competitive or cooperative scenarios.