Stable Baselines3

This is your go-to for production RL if you want battle-tested algorithms without reinventing the wheel. You get PPO, SAC, DQN, TD3, DDPG, and A2C with a scikit-learn style API that actually works. The skill covers the full workflow: training agents, building custom Gymnasium environments, vectorizing for parallel rollouts, and using callbacks for checkpointing and early stopping. It's built on PyTorch 2.3+ and designed for single-agent scenarios. The documentation is clear about constraints, like how SB3 doesn't support Discrete spaces with non-zero starts and normalizes uint8 images automatically. If you need massive parallel training or multi-agent systems, you'll want pufferlib instead, but for standard RL experiments and quick prototyping, this is the reliable choice.