torchrl-agents 0.2.2

Object-oriented RL agents based on torchrl, together with examples on benchmark environments.

TorchRL Agents

TorchRL Agents is a Python package that provides object-oriented reinforcement learning (RL) agents built on top of PyTorch and TorchRL.

Object-Oriented Design

At the core of this package is the Agent abstract base class, which defines a structured interface for RL agents. The Agent class provides a blueprint that other classes can subclass to implement specific RL algorithms. This design ensures consistency and reusability across different agents.

Key Features of the Agent Class:

• Abstract Methods: Enforces the implementation of essential functionalities, such as processing batches and defining policies.
• Serialization: Supports saving and loading agent configurations and weights, enabling easy training and deployment.
• Modularity: Allows for easy extension and customization by subclassing.

By subclassing the Agent class, you can implement various RL algorithms while adhering to a consistent structure.

Examples

The examples folder contains practical demonstrations of how to use the agents implemented in this package. These examples showcase:

• Training RL agents on benchmark environments such as CartPole and Pendulum.
• Using the modular components of the Agent class to customize behavior.
• Saving and loading trained agents for evaluation or deployment.

Getting Started

1. Install uv

2. Clone the repository and create a virtual environment:

   git clone https://github.com/valterschutz/torchrl-examples.git
   uv sync
   

3. Explore the examples folder to see how to train and evaluate RL agents. The scripts can be run without any arguments.

License

This project is licensed under the MIT License. See the LICENSE file for details.