Documentation | Implemented Algorithms | Installation | Getting Started | License

---

OmniSafe is an infrastructural framework designed to accelerate safe reinforcement learning (RL) research. It provides a comprehensive and reliable benchmark for safe RL algorithms, and also an out-of-box modular toolkit for researchers. SafeRL intends to develop algorithms that minimize the risk of unintended harm or unsafe behavior.

OmniSafe stands as the inaugural unified learning framework in the realm of safe reinforcement learning, aiming to foster the Growth of SafeRL Learning Community. The key features of OmniSafe:

• Highly Modular Framework. OmniSafe presents a highly modular framework, incorporating an extensive collection of tens of algorithms tailored for safe reinforcement learning across diverse domains. This framework is versatile due to its abstraction of various algorithm types and well-designed API, using the Adapter and Wrapper design components to bridge gaps and enable seamless interactions between different components. This design allows for easy extension and customization, making it a powerful tool for developers working with different types of algorithms.

• High-performance parallel computing acceleration. By harnessing the capabilities of torch.distributed, OmniSafe accelerates the learning process of algorithms with process parallelism. This enables OmniSafe not only to support environment-level asynchronous parallelism but also incorporates agent asynchronous learning. This methodology bolsters training stability and expedites the training process via the deployment of a parallel exploration mechanism. The integration of agent asynchronous learning in OmniSafe underscores its commitment to providing a versatile and robust platform for advancing SafeRL research.

• Out-of-box toolkits. OmniSafe offers customizable toolkits for tasks like training, benchmarking, analyzing, and rendering. Tutorials and user-friendly APIs make it easy for beginners and average users, while advanced researchers can enhance their efficiency without complex code.

---

Table of Contents

• Quick Start
  • Installation
    • Prerequisites
    • Install from source
    • Install from PyPI
• Implemented Algorithms
  • Examples
    • Algorithms Registry
    • Supported Environments
    • Try with CLI
• Getting Started
  • Important Hints
  • Quickstart: Colab on the Cloud
• Changelog
• Citing OmniSafe
• The OmniSafe Team
• License

---

Quick Start

Installation

Prerequisites

OmniSafe requires Python 3.8+ and PyTorch 1.10+.

We support and test for Python 3.8, 3.9, 3.10 on Linux. Meanwhile, we also support M1 and M2 versions of macOS. We will accept PRs related to Windows, but do not officially support it.

Install from source

# Clone the repo
git clone https://github.com/PKU-Alignment/omnisafe.git
cd omnisafe

# Create a conda environment
conda env create --file conda-recipe.yaml
conda activate omnisafe

# Install omnisafe
pip install -e .

Install from PyPI

OmniSafe is hosted in / .

pip install omnisafe

Implemented Algorithms

Latest SafeRL Papers

• [AAAI 2023] Augmented Proximal Policy Optimization for Safe Reinforcement Learning (APPO)
• [NeurIPS 2022] Constrained Update Projection Approach to Safe Policy Optimization (CUP)
• [NeurIPS 2022] Effects of Safety State Augmentation on Safe Exploration (Simmer)
• [NeurIPS 2022] Model-based Safe Deep Reinforcement Learning via a Constrained Proximal Policy Optimization Algorithm
• [ICML 2022] Sauté RL: Almost Surely Safe Reinforcement Learning Using State Augmentation (SauteRL)
• [IJCAI 2022] Penalized Proximal Policy Optimization for Safe Reinforcement Learning
• [AAAI 2022] Conservative and Adaptive Penalty for Model-Based Safe Reinforcement Learning (CAP)

List of Algorithms

On Policy SafeRL

• The Lagrange version of PPO (PPO-Lag)
• The Lagrange version of TRPO (TRPO-Lag)
• [ICML 2017] Constrained Policy Optimization (CPO)
• [ICLR 2019] Reward Constrained Policy Optimization (RCPO)
• [ICML 2020] Responsive Safety in Reinforcement Learning by PID Lagrangian Methods (PID-Lag)
• [NeurIPS 2020] First Order Constrained Optimization in Policy Space (FOCOPS)
• [AAAI 2020] IPO: Interior-point Policy Optimization under Constraints (IPO)
• [ICLR 2020] Projection-Based Constrained Policy Optimization (PCPO)
• [ICML 2021] CRPO: A New Approach for Safe Reinforcement Learning with Convergence Guarantee
• [IJCAI 2022] Penalized Proximal Policy Optimization for Safe Reinforcement Learning(P3O)

Off Policy SafeRL

• [Preprint 2019] The Lagrangian version of DDPG (DDPGLag)
• [Preprint 2019] The Lagrangian version of TD3 (TD3Lag)
• [Preprint 2019] The Lagrangian version of SAC (SACLag)
• [ICML 2020] Responsive Safety in Reinforcement Learning by PID Lagrangian Methods (DDPGPID)
• [ICML 2020] Responsive Safety in Reinforcement Learning by PID Lagrangian Methods (TD3PID)
• [ICML 2020] Responsive Safety in Reinforcement Learning by PID Lagrangian Methods (SACPID)

Model-Based SafeRL

• [NeurIPS 2021] Safe Reinforcement Learning by Imagining the Near Future (SMBPO)
• [CoRL 2021 (Oral)] Learning Off-Policy with Online Planning (SafeLOOP)
• [AAAI 2022] Conservative and Adaptive Penalty for Model-Based Safe Reinforcement Learning (CAP)
• [NeurIPS 2022] Model-based Safe Deep Reinforcement Learning via a Constrained Proximal Policy Optimization Algorithm
• [ICLR 2022] Constrained Policy Optimization via Bayesian World Models (LA-MBDA)
• [ICML 2022 Workshop] Constrained Model-based Reinforcement Learning with Robust Cross-Entropy Method (RCE)
• [NeurIPS 2018] Constrained Cross-Entropy Method for Safe Reinforcement Learning (CCE)

Offline SafeRL

• The Lagrange version of BCQ (BCQ-Lag)
• The Constrained version of CRR (C-CRR)
• [AAAI 2022] Constraints Penalized Q-learning for Safe Offline Reinforcement Learning CPQ
• [ICLR 2022 (Spotlight)] COptiDICE: Offline Constrained Reinforcement Learning via Stationary Distribution Correction Estimation
• [ICML 2022] Constrained Offline Policy Optimization (COPO)

Others

• [RA-L 2021] Recovery RL: Safe Reinforcement Learning with Learned Recovery Zones
• [ICML 2022] Sauté RL: Almost Surely Safe Reinforcement Learning Using State Augmentation (SauteRL)
• [NeurIPS 2022] Effects of Safety State Augmentation on Safe Exploration

---

Examples

cd examples
python train_policy.py --algo PPOLag --env-id SafetyPointGoal1-v0 --parallel 1 --total-steps 10000000 --device cpu --vector-env-nums 1 --torch-threads 1

Algorithms Registry

Domains	Types	Algorithms Registry
On Policy	Primal Dual	TRPOLag; PPOLag; PDO; RCPO
		TRPOPID; CPPOPID
	Convex Optimization	CPO; PCPO; FOCOPS; CUP
	Penalty Function	IPO; P3O
	Primal	OnCRPO
Off Policy	Primal-Dual	DDPGLag; TD3Lag; SACLag
		DDPGPID; TD3PID; SACPID
Model-based	Online Plan	SafeLOOP; CCEPETS; RCEPETS
	Pessimistic Estimate	CAPPETS
Offline	Q-Learning Based	BCQLag; C-CRR
	DICE Based	COptDICE
Other Formulation MDP	ET-MDP	PPOEarlyTerminated; TRPOEarlyTerminated
	SauteRL	PPOSaute; TRPOSaute
	SimmerRL	PPOSimmerPID; TRPOSimmerPID

Supported Environments

Here is a list of environments that Safety-Gymnasium supports:

Category	Task	Agent	Example
Safe Navigation	Goal[012]	Point, Car, Racecar, Ant	SafetyPointGoal1-v0
	Button[012]
	Push[012]
	Circle[012]
Safe Velocity	Velocity	HalfCheetah, Hopper, Swimmer, Walker2d, Ant, Humanoid	SafetyHumanoidVelocity-v1

For more information about environments, please refer to Safety-Gymnasium.

Customizing your environment

We provide interfaces for customizing environments in the omnisafe/envs directory. You can refer to the examples provided in omnisafe/envs/safety_gymnasium_env to customize the environment interface. Key steps include:

• New a file based on your custom environment, e.g. omnisafe/envs/custom_env.py
• Define the class based on your custom environment, e.g. CustomEnv
• Add comments env_register above the class name to register the environment.

@env_register
class CustomEnv(CMDP):

• List your tasks in _support_envs.

_support_envs: ClassVar[list[str]] = [
      'Custom0-v0',
      'Custom1-v0',
      'Custom2-v0',
    ]

• Redefine self._env in the __init__ function.

self._env = custom_env.make(env_id=env_id, **kwargs)

Next, refer to the SafetyGymnasiumEnv in omnisafe/envs/safety_gymnasium_env to define the step, reset and other functions. Make sure the number, type, order of the returned values match the examples we provided to complete the environment interface design.

Finally, you can run

cd examples
python train_policy.py --algo PPOLag --env Custom1-v0

to run PPOLag in Custom1-v0, as you have registered Custom1-v0 in _support_envs.

Note: If you find trouble customizing your environment, please feel free to open an issue or discussion. Pull requests are also welcomed if you're willing to contribute the implementation of your environments interface.

Try with CLI

pip install omnisafe

omnisafe --help  # Ask for help

omnisafe benchmark --help  # The benchmark also can be replaced with 'eval', 'train', 'train-config'

# Quick benchmarking for your research, just specify:
# 1. exp_name
# 2. num_pool(how much processes are concurrent)
# 3. path of the config file (refer to omnisafe/examples/benchmarks for format)

# Here we provide an exampe in ./tests/saved_source.
# And you can set your benchmark_config.yaml by following it
omnisafe benchmark test_benchmark 2 ./tests/saved_source/benchmark_config.yaml

# Quick evaluating and rendering your trained policy, just specify:
# 1. path of algorithm which you trained
omnisafe eval ./tests/saved_source/PPO-{SafetyPointGoal1-v0} --num-episode 1

# Quick training some algorithms to validate your thoughts
# Note: use `key1:key2`, your can select key of hyperparameters which are recursively contained, and use `--custom-cfgs`, you can add custom cfgs via CLI
omnisafe train --algo PPO --total-steps 2048 --vector-env-nums 1 --custom-cfgs algo_cfgs:steps_per_epoch --custom-cfgs 1024

# Quick training some algorithms via a saved config file, the format is as same as default format
omnisafe train-config ./tests/saved_source/train_config.yaml

---

Getting Started

Important Hints

We have provided benchmark results for various algorithms, including on-policy, off-policy, model-based, and offline approaches, along with parameter tuning analysis. Please refer to the following:

• On-Policy
• Off-Policy
• Model-based
• Offline

Quickstart: Colab on the Cloud

Explore OmniSafe easily and quickly through a series of Google Colab notebooks:

• Getting Started Introduce the basic usage of OmniSafe so that users can quickly hand it.
• CLI Command Introduce how to use the CLI tool of OmniSafe.

We take great pleasure in collaborating with our users to create tutorials in various languages. Please refer to our list of currently supported languages. If you are interested in translating the tutorial into a new language or improving an existing version, kindly submit a PR to us.

---

Changelog

See CHANGELOG.md.

Citing OmniSafe

If you find OmniSafe useful or use OmniSafe in your research, please cite it in your publications.

@article{omnisafe,
  title   = {OmniSafe: An Infrastructure for Accelerating Safe Reinforcement Learning Research},
  author  = {Jiaming Ji, Jiayi Zhou, Borong Zhang, Juntao Dai, Xuehai Pan, Ruiyang Sun, Weidong Huang, Yiran Geng, Mickel Liu, Yaodong Yang},
  journal = {arXiv preprint arXiv:2305.09304},
  year    = {2023}
}

The OmniSafe Team

OmniSafe is mainly developed by the SafeRL research team directed by Prof. Yaodong Yang. Our SafeRL research team members include Borong Zhang, Jiayi Zhou, JTao Dai, Weidong Huang, Ruiyang Sun, Xuehai Pan and Jiaming Ji. If you have any questions in the process of using OmniSafe, don't hesitate to ask your questions on the GitHub issue page, we will reply to you in 2-3 working days.

License

OmniSafe is released under Apache License 2.0.