Algorithms for Safe Reinforcement Learning Problems that were tested and benchmarked in the Bullet-Safety-Gym.

Install this repository with:

git clone https://github.com/SvenGronauer/RL-Safety-Algorithms.git

cd RL-Safety-Algorithms

pip install -e .

Works with every environment that is compatible with the OpenAI Gym interface:

python -m rl_safety_algorithms.train --alg trpo --env MountainCarContinuous-v0

For an open-source framework to benchmark and test safety, we recommend the Bullet-Safety-Gym. To train an algorithms such as Constrained Policy Optimization, run:

python -m rl_safety_algorithms.train --alg cpo --env SafetyBallCircle-v0

In order to benchmark tasks from the Bullet-Safety-Gym, we have prepared scripts in the experiments directory.

cd experiments/
python benchmark_circle_tasks.py

In our experiments, we used a Threadripper 3990X CPU with 64 physical CPU cores, thus, we ran the experiments with the following flag for optimal MPI usage:

python benchmark_circle_tasks.py --num-cores 64

Plots from experiment runs can be also taken from the Bullet-Safety-Gym Benchmarks