// the find
DLR-RM/stable-baselines3
PyTorch version of Stable Baselines, reliable implementations of reinforcement learning algorithms.
Stable Baselines3 is the standard PyTorch implementation of classic RL algorithms (PPO, SAC, TD3, DQN, A2C, DDPG, HER), maintained by the researchers who wrote the original. It exists to give the ML community clean, tested, benchmarked baselines rather than forcing everyone to reimplement the same algorithms slightly differently. If you're doing applied RL research or building a project that needs a solid PPO/SAC starting point, this is the one to use.
The implementations are actually verified against published results — there are benchmark logs on OpenRL/W&B and the code is tested against known-good performance numbers, not just unit tested for correctness. The sklearn-style API (model.learn(), model.predict()) means you can swap algorithms without restructuring your code. Full type hints throughout, mypy passing, ruff for linting — the code quality is unusually high for an academic project. The ecosystem split is smart: stable core in SB3, experimental stuff in sb3-contrib, JAX variants in SBX — so the main library stays auditable.
No recurrent policies in the core library — LSTM/RecurrentPPO lives in sb3-contrib, which means if you need memory you're one extra dependency away from the stable release cycle. The vectorized environment abstraction (VecEnv) predates the current Gymnasium API and the shim layer (patch_gym.py) is a sign that the architecture is showing its age. Multi-agent RL is completely out of scope — this is single-agent only, so if your problem involves multiple interacting agents you're hitting a wall immediately. The JAX version (SBX) is faster by up to 20x but is a separate project with a fraction of the features, so there's no clean upgrade path if you need speed later.