stable-baselines3

Train RL agents using Stable Baselines3 by instantiating an algorithm (PPO, SAC, DQN, TD3, A2C, DDPG, or HER), passing a Gymnasium environment, and calling the learn() method. SB3 provides a unified API across all algorithms, handles vectorized environments automatically, and supports model checkpointing during training.

Yes, Stable Baselines3 works with any custom Gymnasium-compliant environment. Design your environment following the Gymnasium API, validate it with the provided checker, then pass it directly to any SB3 algorithm. Custom environments enable domain-specific reward structures and state spaces.

Stable Baselines3 implements PPO, SAC, TD3, DQN, A2C, DDPG, and HER algorithms. Each algorithm has battle-tested, production-ready implementations with unified interfaces, making it straightforward to benchmark different approaches on the same environment without rewriting training loops.

Stable Baselines3 supports callbacks for real-time monitoring and evaluation routines that track agent performance during and after training. Callbacks integrate with vectorized environments to log metrics, save checkpoints, and halt training based on performance thresholds without disrupting the training pipeline.

Yes, Stable Baselines3 natively supports vectorized environments for parallel agent training across multiple environment instances. Vectorization accelerates data collection and training on multi-core systems, scaling sample efficiency without algorithm changes or custom parallelization code.

Stable Baselines3 provides built-in model saving and loading functions that preserve trained agent weights and hyperparameters. Load a saved model to continue training, evaluate on new environments, or deploy in production—all without retraining or manual parameter management.