A comprehensive and reliable benchmark for safe reinforcement learning.
Reason this release was yanked:
Some inevitable bugs appear in this v0.2.1.
Project description
Documentation | Implemented Algorithms | Installation | Getting Started | License
This library is currently under heavy development - if you have suggestions on the API or use-cases you'd like to be covered, please open a GitHub issue or reach out. We'd love to hear about how you're using the library.
OmniSafe is an infrastructural framework designed to accelerate safe reinforcement learning (RL) research by providing a comprehensive and reliable benchmark for safe RL algorithms. The field of RL has great potential to benefit society, but safety concerns are a significant issue, and RL algorithms have raised concerns about unintended harm or unsafe behavior. Safe RL intends to develop algorithms that minimize the risk of unintended harm or unsafe behavior, but there is currently a lack of commonly recognized safe RL algorithm benchmarks.
OmniSafe addresses these issues by providing more than 40 experimentally validated algorithms and a sound and efficient simulation environment. Researchers can use OmniSafe to conduct experiments and verify their ideas, ensuring consistency and enabling more efficient development of safe RL algorithms. By using OmniSafe as a benchmark, researchers can evaluate the performance of their own safe RL algorithms and contribute to the advancement of safe RL research.
Table of Contents
Implemented Algorithms
The supported interface algorithms currently include:
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)
- [ICML 2022] Constrained Variational Policy Optimization for Safe Reinforcement Learning (CVPO)
- [IJCAI 2022] Penalized Proximal Policy Optimization for Safe Reinforcement Learning
- [ICLR 2022] Constrained Policy Optimization via Bayesian World Models (LA-MBDA)
- [AAAI 2022] Conservative and Adaptive Penalty for Model-Based Safe Reinforcement Learning (CAP)
List of Algorithms
On-Policy Safe
- 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 Safe
- The Lagrange version of TD3 (TD3-Lag)
- The Lagrange version of DDPG (DDPG-Lag)
- The Lagrange version of SAC (SAC-Lag)
- [ICML 2019] Lyapunov-based Safe Policy Optimization for Continuous Control (SDDPG)
- [ICML 2019] Lyapunov-based Safe Policy Optimization for Continuous Control (SDDPG-modular)
- [ICML 2022] Constrained Variational Policy Optimization for Safe Reinforcement Learning (CVPO)
Model-Based Safe
- [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)
Offline Safe
- 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
- Safe Exploration in Continuous Action Spaces (Safety Layer)
- [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
Installation
Prerequisites
OmniSafe requires Python 3.8+ and PyTorch 1.10+.
Install from source
# Clone the repo
git clone https://github.com/PKU-MARL/omnisafe
cd omnisafe
# Create a conda environment
conda create -n omnisafe python=3.8
conda activate omnisafe
# Install omnisafe
pip install -e .
Install from PyPI
pip install omnisafe
Examples
cd examples
python train_policy.py --algo PPOLag --env-id SafetyPointGoal1-v0 --parallel 1 --total-steps 1024000 --device cpu --vector-env-nums 1 --torch-threads 1
algo:
Type | Name |
---|---|
Base-On-Policy |
PolicyGradient, PPO NaturalPG, TRPO |
Base-Off-Policy |
DDPG, TD3, SAC |
Naive Lagrange |
RCPO, PPOLag, TRPOLag DDPGLag, TD3Lag, SACLag |
PID Lagrange |
CPPOPid, TRPOPid |
First Order |
FOCOPS, CUP |
Second Order |
SDDPG, CPO, PCPO |
Saute RL |
PPOSaute, PPOLagSaute |
Simmer RL |
PPOSimmerQ, PPOSimmerPid PPOLagSimmerQ, PPOLagSimmerPid |
EarlyTerminated |
PPOEarlyTerminated PPOLagEarlyTerminated |
Model-Based |
CAP, MBPPOLag, SafeLOOP |
env-id: Environment id in Safety Gymnasium, here a list of envs 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-v4 |
More information about environments, please refer to Safety Gymnasium
parallel: Number of parallels
Try with CLI
A video example
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)
omnisafe benchmark test_benchmark 2 ./saved_source/benchmark_config.yaml
# Quick evaluating and rendering your trained policy, just specify: 1.path of algorithm which you trained
omnisafe eval ./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 1024 --vector-env-nums 1 --custom-cfgs algo_cfgs:update_cycle --custom-cfgs 512
# Quick training some algorithms via a saved config file, the format is as same as default format
omnisafe train-config ./saved_source/train_config.yaml
Getting Started
Important Hints
train_cfgs:torch_threads
is especially important for training speed, and is varying with users' machine, this value shouldn't be too small or too large.
1. Run Agent from preset yaml file
import omnisafe
env_id = 'SafetyPointGoal1-v0'
agent = omnisafe.Agent('PPOLag', env_id)
agent.learn()
2. Run agent with custom cfg
import omnisafe
env_id = 'SafetyPointGoal1-v0'
custom_cfgs = {
'train_cfgs': {
'total_steps': 1024000,
'vector_env_nums': 1,
'parallel': 1,
},
'algo_cfgs': {
'update_cycle': 2048,
'update_iters': 1,
},
'logger_cfgs': {
'use_wandb': False,
},
}
agent = omnisafe.Agent('PPOLag', env_id, custom_cfgs=custom_cfgs)
agent.learn()
3. Run Agent from custom terminal config
You can also run agent from a custom terminal config. You can set any config in a corresponding yaml file.
For example, you can run PPOLag
agent on SafetyPointGoal1-v0
environment with total_steps=1024000
, vector_env_nums=1
and parallel=1
by:
cd examples
python train_policy.py --algo PPOLag --env-id SafetyPointGoal1-v0 --parallel 1 --total-steps 1024000 --device cpu --vector-env-nums 1 --torch-threads 1
4. Evalutate Saved Policy
import os
import omnisafe
# Just fill your experiment's log directory in here.
# Such as: ~/omnisafe/runs/SafetyPointGoal1-v0/CPO/seed-000-2022-12-25_14-45-05
LOG_DIR = ''
evaluator = omnisafe.Evaluator()
for item in os.scandir(os.path.join(LOG_DIR, 'torch_save')):
if item.is_file() and item.name.split('.')[-1] == 'pt':
evaluator.load_saved_model(save_dir=LOG_DIR, model_name=item.name)
evaluator.render(num_episode=10, camera_name='track', width=256, height=256)
Changelog
See CHANGELOG.md.
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.
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distribution
Hashes for omnisafe-0.2.1-py2.py3-none-any.whl
Algorithm | Hash digest | |
---|---|---|
SHA256 | 489002c475cc5215727da564039430e3cd4557d45e725d3f0bf117b68e8a4fed |
|
MD5 | 9fc39cc3246d0bdc10eacb08c24d7433 |
|
BLAKE2b-256 | 7b22fd401252b356d4bdffcf36c8ea9826fcfc67766355f29520e03b3b63229c |