Skip to content

Installation

safety_sb3 is a pure-Python add-on for Stable-Baselines3. The core install pulls in SB3, PyTorch, Gymnasium and NumPy and nothing exotic.

Not on PyPI

The package is not published on PyPI. pip install safety_sb3 will not find it. Install from the GitHub repository, pinned to a release tag for reproducibility (the current release is v0.4.0).

Supported versions

requirement
Python ≥ 3.10 (validated on 3.10 and 3.11)
Stable-Baselines3 ≥ 2.0.0
PyTorch ≥ 1.13
Gymnasium ≥ 0.28
NumPy ≥ 1.23

The optional benchmark submodules (below) pin Python 3.10 exactly for safety-gymnasium's sake; the core library itself is happy on any ≥ 3.10.

Pin to a release tag so an environment is reproducible:

pip install "safety_sb3 @ git+https://github.com/SafeRoboticsLab/safety-stable-baselines.git@v0.4.0"

To depend on it from another project's requirements, use the same specifier. Pin the tag — v0.4.0 is a breaking rename of every learner class (see the release notes); an unpinned dependency can move under you.

If you need the pre-rename (Isaacs* / Gameplay*) class names, pin the last release that had them:

pip install "safety_sb3 @ git+https://github.com/SafeRoboticsLab/safety-stable-baselines.git@v0.3.0"

Editable / contributor install

Clone and install in editable mode to hack on the library or run the examples:

conda create --name safety_sb3 python=3.10   # 3.10 keeps the optional submodules happy
conda activate safety_sb3

git clone https://github.com/SafeRoboticsLab/safety-stable-baselines.git
cd safety-stable-baselines
pip install -e .
pip install -e ".[tests]"       # adds pytest, to run the test suite

Optional: benchmark / integration dependencies

Only needed to run the bundled examples/ against the external benchmark environments. This repo vendors safety-gymnasium (SafeRoboticsLab fork) and rl_baselines3_zoo as git submodules (Python 3.10 is pinned for safety-gymnasium):

git submodule update --init
pip install -e integrations/rl_baselines3_zoo
pip install -e integrations/safety-gymnasium

The shipped reference environment (Bicycle5D) and the Pendulum quickstart need none of these — they are NumPy-only and CPU-trainable.

GPU considerations

The core library trains fine on CPU (the Quickstart and Bicycle5D converge in minutes on a laptop). A CUDA build of PyTorch is only worth it for the tensor path — thousands of parallel, GPU-resident environments via TensorVecEnv. Install the appropriate PyTorch CUDA wheel for your driver before installing safety_sb3 if you need it; nothing in this package constrains the CUDA version.

Verify the install

python -c "import safety_sb3; print('installation OK')"

For a stronger check that the learner classes import cleanly:

python -c "from safety_sb3 import SafetySAC1P, ReachAvoidPPO1P, SafetySAC2P; print('learners OK')"

Then run the test suite (editable install):

python -m pytest tests/ -q

Next steps