This repository provides a PyTorch implementation of various model-free and model-based deep reinforcement learning algorithms, targeting researchers and practitioners working with continuous action spaces and OpenAI Gym environments. It aims to simplify the process of training and evaluating state-of-the-art RL algorithms.

How It Works

The library implements algorithms directly in PyTorch, allowing for efficient training on both CPU and GPU. It integrates seamlessly with OpenAI Gym, enabling easy experimentation across a wide range of environments, including robotics tasks and Atari games. The project also explores advanced techniques like Hindsight Experience Replay (HER) and Prioritized Experience Replay, and incorporates research ideas for areas like curiosity-driven exploration and world models.

Quick Start & Requirements

• Install via pip: pip install pytorch-policy
• Dependencies: PyTorch, OpenAI Gym, mujoco-py (for specific Gym environments), Tensorboardx.
• Requires Python. GPU acceleration is supported.

Highlighted Details

• Implements algorithms such as DQN, DDPG, PPO, Soft Actor-Critic, and DARLA.
• Supports continuous action spaces and integrates with OpenAI Gym environments (e.g., Fetch, Atari).
• Includes implementations or research on Prioritized Experience Replay, Hindsight Experience Replay, and curiosity-driven exploration.
• Features research on GANs for environment modeling, including InfoGAN and CVAE-GAN with Spectral Normalization.

Maintenance & Community

The repository appears to be a personal project with ongoing research contributions. Links to related TensorFlow implementations and specific algorithm papers are provided. No explicit community channels (Discord, Slack) or roadmap are mentioned.

Licensing & Compatibility

The README does not explicitly state a license. Compatibility for commercial use or closed-source linking is not specified.

Limitations & Caveats

Several advanced algorithms (e.g., Rainbow DQN, A3C, ACER) are listed as "Coming Soon." The project includes "Research" sections, indicating some implementations may be experimental or not fully stable. The README also notes the difficulty in training GANs even with advanced techniques.