v2.9.2: Bug fixes and doc improvements

Warning

Stable-Baselines3 (SB3) v2.9.2 will be the last one supporting Python 3.10 (end of life in October 2026)
We highly recommended you to upgrade to Python >= 3.11.

SB3 Contrib (more algorithms): https://github.com/Stable-Baselines-Team/stable-baselines3-contrib
RL Zoo3 (training framework): https://github.com/DLR-RM/rl-baselines3-zoo
Stable-Baselines Jax (SBX): https://github.com/araffin/sbx

To upgrade:

pip install stable_baselines3 sb3_contrib rl_zoo3 --upgrade

Bug Fixes:

• Fixed file descriptor leak in SubprocVecEnv.close() by closing parent-side pipes to prevent "too many open files" errors in long-running loops
• Fixed Pandas deprecation warnings on inplace operations

SB3-Contrib

• Upgraded to Stable-Baselines3 >= 2.9.2

RL Zoo

• Set learning_starts to 500 for parking-v0 in TQC hyperparameters to ensure the first episode is finished
• Fixed "too many open files" issue when doing hyperparameter optimization: the trial environments are now closed in a finally block and gc_after_trial=True is passed to Optuna to free them between trials (also needs latest SB3 master version to fully work)
• Set use_expln=True for pybullet envs when using gSDE to avoid NaNs
• Add a note regarding pybullet source installation error on macOS due to legacy code
• Migrate packaging setup from setup.py to pyproject.toml and integrate requirements.txt as a new extra

SBX (SB3 + Jax)

• Updated Jax dependency version range

Deprecations:

Others:

• Fixed moviepy test compatibility with moviepy >= 2.0 (PyTorch doesn't support it yet)
• Added filter for numpy newshape deprecation warning from torch tensorboard
• Fixed mypy error in test_vec_envs.py by wrapping itertools.product with list()
• Improved tests coverage for RMSpropTFLike optimizer (invalid params, centered/momentum/weight_decay branches, pickle roundtrip)
• Improved test coverage for save_util.py: added tests for BadZipFile error handling in load_from_zip_file and IsADirectoryError/FileNotFoundError handling in open_path
• Fixed Docker build by adding the missing --system flag to uv pip uninstall opencv-python (required by recent uv)
• Added a warning when n_steps > 1 is passed together with a custom replay_buffer_class, as n_steps is only applied when the replay buffer class is selected automatically (@Koustav-github)
• Updated tests dependencies

Documentation:

• Clarified in the n_steps docstring that it only applies when replay_buffer_class is None, and that a custom replay buffer must be configured with replay_buffer_kwargs={"n_steps": ..., "gamma": ...} (@Koustav-github)
• Updated algorithm table (added the ones from SBX and sorted the table)
• Fixed typos, grammar, and Gymnasium API inconsistencies across documentation

New Contributors

• @nikolauspschuetz made their first contribution in #2272
• @Koustav-github made their first contribution in #2275

Full Changelog: v2.9.0...v2.9.2