RLzoo is a comprehensive reinforcement learning library designed for researchers and practitioners to easily implement, benchmark, and develop RL algorithms. It offers a wide range of popular RL algorithms and supports various environments, including OpenAI Gym, DeepMind Control Suite, and RLBench, with plans for future support of larger-scale distributed training frameworks.

How It Works

RLzoo leverages TensorFlow 2.0 and TensorLayer 2.0+ for its neural network implementations. It provides a flexible API that allows users to configure algorithms and environments either implicitly through default configuration files or explicitly within their scripts. This design promotes interpretability and ease of use for both new learners and experienced researchers.

Quick Start & Requirements

• Installation: pip3 install rlzoo --upgrade or git clone and pip3 install .
• Prerequisites: TensorFlow >= 2.0.0, TensorLayer >= 2.0.1, TensorFlow-Probability. Additional dependencies like Mujoco, dm_control, or Vrep/PyRep/RLBench are required for specific environments.
• Usage: Run python run_rlzoo.py from the root directory for a quick start. Detailed examples and interactive configurations via Jupyter Notebook are available.
• Documentation: Online Documentation

Highlighted Details

• Supports a wide array of RL algorithms including DQN, PPO, SAC, TD3, DDPG, and more.
• Integrates with diverse environments: Atari, Box2D, Classic Control, MuJoCo, Robotics, DeepMind Control Suite, and RLBench.
• Offers automatic model construction adapting to observation and action spaces (discrete/continuous, pixel/vector).
• Includes support for distributed training using the Kungfu package.

Maintenance & Community

The project is actively seeking community contributions. Discussions and bug reporting are encouraged via GitHub issues and a Slack channel.

Licensing & Compatibility

The repository does not explicitly state a license in the provided README. Users should verify licensing for commercial or closed-source integration.

Limitations & Caveats

Default hyperparameters may not be optimal for all environments and algorithms. Training with raw-pixel observations can be challenging and may require extensive hyperparameter tuning. The README mentions potential issues in the coming months after initial release, indicating ongoing development.