Simple Playgrounds provides a modular platform for building interactive 2D grid environments where agents can navigate mazes, interact with objects, and complete tasks. Users define environment layouts, object behaviors, and reward functions via simple YAML or Python scripts. The integrated Pygame renderer delivers real-time visualization, while a step-based API ensures seamless integration with reinforcement learning libraries like Stable Baselines3. With support for multi-agent setups, collision detection, and customizable physics parameters, Simple Playgrounds streamlines the prototyping, benchmarking, and educational demonstration of AI algorithms.

RL-Agents is a research-grade reinforcement learning framework built on PyTorch that bundles popular RL algorithms across value-based, policy-based, and actor-critic methods. The library features a modular agent API, GPU acceleration, seamless integration with OpenAI Gym, and built-in logging and visualization tools. Users can configure hyperparameters, customize training loops, and benchmark performance with a few lines of code, making RL-Agents ideal for academic research, prototyping, and industrial experimentation.

RL Collision Avoidance provides a complete pipeline for developing, training, and deploying multi-robot collision avoidance policies. It offers a set of Gym-compatible simulation scenarios where agents learn collision-free navigation through reinforcement learning algorithms. Users can customize environment parameters, leverage GPU acceleration for faster training, and export learned policies. The framework also integrates with ROS for real-world testing, supports pre-trained models for immediate evaluation, and features tools for visualizing agent trajectories and performance metrics.

Dead-simple self-learning offers developers a dead-simple approach to create and train reinforcement learning agents in Python. The framework abstracts core RL components, such as environment wrappers, policy modules, and experience buffers, into concise interfaces. Users can quickly initialize environments, define custom policies using familiar PyTorch or TensorFlow backends, and execute training loops with built-in logging and checkpointing. The library supports on-policy and off-policy algorithms, enabling flexible experimentation with Q-learning, policy gradients, and actor-critic methods. By reducing boilerplate code, dead-simple self-learning allows practitioners, educators, and researchers to prototype algorithms, test hypotheses, and visualize agent performance with minimal configuration. Its modular design also facilitates integration with existing ML stacks and custom environments.

SoccerAgent is a specialized AI framework designed for developing and training autonomous soccer agents using state-of-the-art multi-agent reinforcement learning (MARL) techniques. It simulates realistic soccer matches in 2D or 3D environments, offering tools to define reward functions, customize player attributes, and implement tactical strategies. Users can integrate popular RL algorithms (such as PPO, DDPG, and MADDPG) via built-in modules, monitor training progress through dashboards, and visualize agent behaviors in real time. The framework supports scenario-based training for offense, defense, and coordination protocols. With an extensible codebase and detailed documentation, SoccerAgent empowers researchers and developers to analyze team dynamics and refine AI-driven gameplay strategies for academic and commercial projects.

This repository provides an end-to-end reinforcement learning framework for StarCraft II gameplay research. The core agent uses Proximal Policy Optimization (PPO) to learn policy networks that interpret observation data from the PySC2 environment and output precise in-game actions. Developers can configure neural network layers, reward shaping, and training schedules to optimize performance. The system supports multiprocessing for efficient sample collection, logging utilities for monitoring training curves, and evaluation scripts for running trained policies against scripted or built-in AI opponents. The codebase is written in Python and leverages TensorFlow for model definition and optimization. Users can extend components such as custom reward functions, state preprocessing, or network architectures to suit specific research objectives.

TexasHoldemAgent provides a modular environment built on Python to train, evaluate, and deploy an AI-powered poker player for heads-up limit Texas Hold’em. It integrates a custom simulation engine with deep reinforcement learning algorithms, including DQN, for iterative policy improvement. Key capabilities include hand state encoding, action space definition (fold, call, raise), reward shaping, and real-time decision evaluation. Users can customize learning parameters, leverage CPU/GPU acceleration, monitor training progress, and load or save trained models. The framework supports batch simulation to test various strategies, generate performance metrics, and visualize win rates, empowering researchers, developers, and poker enthusiasts to experiment with AI-driven gameplay strategies.

Text-to-Reward provides a pipeline to train reward models that map text-based task descriptions or feedback into scalar reward values for RL agents. Leveraging transformer-based architectures and fine-tuning on collected human preference data, the framework automatically learns to interpret natural language instructions as reward signals. Users can define arbitrary tasks via text prompts, train the model, and then incorporate the learned reward function into any RL algorithm. This approach eliminates manual reward shaping, boosts sample efficiency, and enables agents to follow complex multi-step instructions in simulated or real-world environments.

uAgents is a modular JavaScript framework that empowers developers to build autonomous, decentralized AI agents which can discover peers, exchange messages, collaborate on tasks, and adapt through learning. Agents communicate over libp2p-based gossip protocols, register capabilities via on-chain registries, and negotiate service-level agreements using smart contracts. The core library handles agent lifecycle events, message routing, and extensible behaviors such as reinforcement learning and market-driven task allocation. Through customizable plugins, uAgents can integrate with Fetch.ai’s ledger, external APIs, and oracle networks, enabling agents to perform real-world actions, data acquisition, and decision-making in distributed environments without centralized orchestration.

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.

VMAS is a comprehensive toolkit for building and training multi-agent systems using deep reinforcement learning. It supports GPU-based parallel simulation of hundreds of environment instances, enabling high-throughput data collection and scalable training. VMAS includes implementations of popular MARL algorithms like PPO, MADDPG, QMIX, and COMA, along with modular policy and environment interfaces for rapid prototyping. The framework facilitates centralized training with decentralized execution (CTDE), offers customizable reward shaping, observation spaces, and callback hooks for logging and visualization. With its modular design, VMAS seamlessly integrates with PyTorch models and external environments, making it ideal for research in cooperative, competitive, and mixed-motive tasks across robotics, traffic control, resource allocation, and game AI scenarios.

The YGO-Agent framework allows researchers and enthusiasts to develop AI bots that play the Yu-Gi-Oh card game using reinforcement learning. It wraps the YGOPRO game simulator into an OpenAI Gym-compatible environment, defining state representations such as hand, field, and life points, and action representations including summoning, spell/trap activation, and attacking. Rewards are based on win/loss outcomes, damage dealt, and game progress. The agent architecture uses PyTorch to implement DQN, with options for custom network architectures, experience replay, and epsilon-greedy exploration. Logging modules record training curves, win rates, and detailed move logs for analysis. The framework is modular, enabling users to replace or extend components such as the reward function or action space.

A-Mem is designed to seamlessly integrate with Python-based AI agent frameworks, offering three distinct memory modules: episodic memory for per-episode context, short-term memory for immediate past actions, and long-term memory for accumulating knowledge over time. Developers can customize memory capacity, retention policies, and serialization backends such as in-memory or Redis storage. The library includes efficient indexing algorithms to retrieve relevant memories based on similarity and context windows. By inserting A-Mem’s memory handlers into the agent’s perception-action loop, users can store observations, actions, and outcomes, then query past experiences to inform current decisions. This modular design supports rapid experimentation in reinforcement learning, conversational AI, robotics navigation, and other agent-driven tasks requiring context awareness and temporal reasoning.

GYM_XPLANE_ML wraps the X-Plane flight simulator as an OpenAI Gym environment, exposing throttle, elevator, aileron and rudder controls as action spaces and flight parameters like altitude, speed, and orientation as observations. Users can script training workflows in Python, select predefined scenarios or customize waypoints, weather conditions, and aircraft models. The library handles low-latency communication with X-Plane, runs episodes in synchronous mode, logs performance metrics, and supports real-time rendering for debugging. It enables iterative development of ML-driven autopilots and experimental RL algorithms in a high-fidelity flight environment.

Acme is a Python-based framework that simplifies the development and evaluation of reinforcement learning agents. It offers a collection of prebuilt agent implementations (e.g., DQN, PPO, SAC), environment wrappers, replay buffers, and distributed execution engines. Researchers can mix and match components to prototype new algorithms, monitor training metrics with built-in logging, and leverage scalable distributed pipelines for large-scale experiments. Acme integrates with TensorFlow and JAX, supports custom environments via OpenAI Gym interfaces, and includes utilities for checkpointing, evaluation, and hyperparameter configuration.

AI-Agentic Machine Translation is an open-source framework designed for research and development in machine translation. It orchestrates three core agents—a generator, an evaluator, and a refiner—to collaboratively produce, assess, and refine translations. Built on PyTorch and transformer models, the system supports supervised pre-training, reinforcement learning optimization, and configurable agent policies. Users can benchmark on standard datasets, track BLEU scores, and extend the pipeline with custom agents or reward functions to explore agentic collaboration in translation tasks.

AI Hedge Fund 5zu provides a complete pipeline for quantitative trading: a customizable environment for simulating multiple asset classes, reinforcement learning–based agent modules, backtesting utilities, real-time market data integration, and risk management tools. Users can configure data sources, define reward functions, train agents on historical data, and evaluate performance across key financial metrics. The framework supports modular strategy development and can be extended to live broker APIs for deploying production-level trading bots.

AI Agents for Rock Paper Scissors is an open-source Python project that demonstrates how to build, train, and evaluate different AI strategies—random play, rule-based pattern recognition, and reinforcement learning (Q-learning)—in the classic Rock-Paper-Scissors game. It provides modular agent classes, a configurable game runner, performance logging, and visualization utilities. Users can easily swap agents, adjust learning parameters, and explore AI behavior in competitive scenarios.

Ant_racer is a virtual multi-agent pursuit-evasion platform that provides a game environment for studying multi-agent reinforcement learning. Built on OpenAI Gym and Mujoco, it allows users to simulate interactions between multiple autonomous agents in pursuit and evasion tasks. The platform supports implementation and testing of reinforcement learning algorithms such as DDPG in a physically realistic environment. It is useful for researchers and developers interested in AI multi-agent behaviors in dynamic scenarios.

The Beer Game Environment provides a discrete-time simulation of a four-stage beer supply chain—retailer, wholesaler, distributor, and manufacturer—exposing an OpenAI Gym interface. Agents receive observations including on-hand inventory, pipeline stock, and incoming orders, then output order quantities. The environment computes per-step costs for inventory holding and backorders, and supports customizable demand distributions and lead times. It integrates seamlessly with popular RL libraries like Stable Baselines3, enabling researchers and educators to benchmark and train algorithms on supply chain optimization tasks.