This repository provides a comprehensive training framework for reinforcement learning agents using Stable Baselines3. It targets researchers and practitioners needing to train, evaluate, and benchmark RL algorithms across a wide array of environments, offering pre-tuned hyperparameters and trained agents for accelerated development and reproducible results.

How It Works

The framework leverages a configuration-driven approach, with hyperparameters for various algorithms and environments defined in YAML files. It provides command-line scripts for training, evaluation, hyperparameter tuning, and agent visualization. The design emphasizes modularity, allowing easy integration of new algorithms and environments, and supports experiment tracking via integrations like Weights & Biases and model sharing through Hugging Face.

Quick Start & Requirements

• Install: pip install -e . (from source) or pip install rl_zoo3 (as package).
• Prerequisites: swig, cmake, ffmpeg for full installation. Note: NumPy < 2.0 is required for PyBullet environments.
• Documentation: https://rl-baselines3-zoo.readthedocs.io/
• Colab Notebook: Try it Online!

Highlighted Details

• Over 200 pre-trained agents available for various environments including Atari, Classic Control, Box2D, PyBullet, MuJoCo, Robotics, and MiniGrid.
• Supports multiple RL algorithms: A2C, PPO, DQN, QR-DQN, DDPG, SAC, TD3, TQC, TRPO, ARS, HER.
• Includes scripts for hyperparameter tuning, plotting results, and recording agent videos.
• Integrates with Weights & Biases for experiment tracking and Hugging Face for model sharing.

Maintenance & Community

• Active contributors include @iandanforth, @tatsubori, @Shade5, @mcres, @ernestum, @qgallouedec.
• Aims to expand the collection of trained agents and tuned hyperparameters.

Licensing & Compatibility

• Licensed under the MIT License.
• Compatible with commercial use and closed-source linking.

Limitations & Caveats

• The benchmark results are based on single runs and may not represent maximal performance or statistical significance.
• Compatibility with NumPy >= 2.0 is broken for PyBullet environments due to a dependency issue.