SKRL - Reinforcement Learning library (0.7.0)

skrl is an open-source modular library for Reinforcement Learning written in Python (using PyTorch) and designed with a focus on readability, simplicity, and transparency of algorithm implementation. In addition to supporting the OpenAI Gym and DeepMind environment interfaces, it allows loading and configuring NVIDIA Isaac Gym and NVIDIA Omniverse Isaac Gym environments, enabling agents’ simultaneous training by scopes (subsets of environments among all available environments), which may or may not share resources, in the same run

Main features:

• Clean code

• Modularity and reusability

• Documented library, code and implementations

• Support for OpenAI Gym (single and vectorized), DeepMind, NVIDIA Isaac Gym (preview 2, 3 and 4) and NVIDIA Omniverse Isaac Gym environments

• Simultaneous learning by scopes in OpenAI Gym (vectorized), NVIDIA Isaac Gym and NVIDIA Omniverse Isaac Gym

Warning

skrl is under active continuous development. Make sure you always have the latest version

GitHub repository: https://github.com/Toni-SM/skrl

Questions or discussions: https://github.com/Toni-SM/skrl/discussions

Citing skrl: To cite this library (created at Mondragon Unibertsitatea) use the following reference to its article: “skrl: Modular and Flexible Library for Reinforcement Learning”

@article{serrano2022skrl,
title={skrl: Modular and Flexible Library for Reinforcement Learning},
author={Serrano-Mu{\~n}oz, Antonio and Arana-Arexolaleiba, Nestor and Chrysostomou, Dimitrios and B{\o}gh, Simon},
journal={arXiv preprint arXiv:2202.03825},
year={2022}
}

---

User guide

• Installation
  • Prerequisites
  • Library Installation
  • Troubleshooting
  • Known issues
• Getting Started
  • 1. Environments
  • 2. Memories
  • 3. Models
  • 4. Noises
  • 5. Learning rate schedulers
  • 6. Agents
  • 7. Trainers
• Examples
  • Learning in a Gym environment (one agent, one environment)
  • Learning in a Gym vectorized environment (one agent, multiple environments)
  • Learning in a DeepMind environment (one agent, one environment)
  • Learning in an Isaac Gym environment (one agent, multiple environments)
  • Learning by scopes in an Isaac Gym environment (multiple agents and environments)
  • Learning in an Omniverse Isaac Gym environment (one agent, multiple environments)
  • Learning in an Omniverse Isaac Sim environment (one agent, one environment)
  • Library utilities (skrl.utils module)
• Saving, loading and logging
  • Tracking metrics (TensorBoard)
  • Model checkpoint
  • Memory export/import

---

Library components (overview)

Agents

Definition of reinforcement learning algorithms that compute an optimal policy. All agents inherit from one and only one base class (that defines a uniform interface and provides for common functionalities) but which is not tied to the implementation details of the algorithms

• Advantage Actor Critic (A2C)

• Cross-Entropy Method (CEM)

• Deep Deterministic Policy Gradient (DDPG)

• Double Deep Q-Network (DDQN)

• Deep Q-Network (DQN)

• Proximal Policy Optimization (PPO)

• Q-learning (Q-learning)

• Soft Actor-Critic (SAC)

• State Action Reward State Action (SARSA)

• Twin-Delayed DDPG (TD3)

• Trust Region Policy Optimization (TRPO)

Environments

Definition of the Isaac Gym (preview 2, 3 and 4) and Omniverse Isaac Gym environment loaders, and wrappers for the OpenAI Gym, DeepMind, Isaac Gym and Omniverse Isaac Gym environments

• Wrapping OpenAI Gym, DeepMind, Isaac Gym and Omniverse Isaac Gym environments

• Loading Isaac Gym environments

• Loading Omniverse Isaac Gym environments

Memories

Generic memory definitions. Such memories are not bound to any agent and can be used for any role such as rollout buffer or experience replay memory, for example. All memories inherit from a base class that defines a uniform interface and keeps track (in allocated tensors) of transitions with the environment or other defined data

• Random memory

Models

Definition of helper classes for the construction of tabular functions or function approximators using artificial neural networks. This library does not provide predefined policies but helper classes to create discrete and continuous (stochastic or deterministic) policies in which the user only has to define the tables (tensors) or artificial neural networks. All models inherit from one base class that defines a uniform interface and provides for common functionalities

• Tabular model (discrete domain)

• Categorical model (discrete domain)

• Gaussian model (continuous domain)

• Deterministic model (continuous domain)

Trainers

Definition of the procedures responsible for managing the agent’s training and interaction with the environment. All trainers inherit from a base class that defines a uniform interface and provides for common functionalities

• Sequential trainer

• Parallel trainer

Resources

Definition of resources used by the agents during training and/or evaluation, such as exploration noises or learning rate schedulers

Noises: Definition of the noises used by the agents during the exploration stage. All noises inherit from a base class that defines a uniform interface

• Gaussian noise

• Ornstein-Uhlenbeck noise

Learning rate schedulers: Definition of learning rate schedulers. All schedulers inherit from the PyTorch _LRScheduler class (see how to adjust learning rate in the PyTorch documentation for more details)

• KL Adaptive

Preprocessors: Definition of preprocessors

• Running standard scaler

Utils

Definition of helper functions and classes

• Utilities, e.g. setting the random seed

• Model instantiators

• Memory and Tensorboard file post-processing

• Isaac Gym utils