stable-baselines3

Training RL agents with Stable Baselines3 involves setting up a Gymnasium-compatible environment, selecting an algorithm (PPO, SAC, DQN, TD3, A2C, etc.), and calling the learn() method on your agent. SB3 handles the training loop, policy updates, and gradient computation across your chosen algorithm, returning a trained model ready for evaluation or deployment.

Vectorized environments in Stable Baselines3 run multiple environment instances in parallel, collecting experience faster without sequential overhead. Use VecEnv wrappers to batch environment steps, then pass the vectorized wrapper to your agent; this scales data collection linearly with the number of parallel environments, cutting wall-clock training time significantly.

Yes. Stable Baselines3 provides EvalCallback to periodically test your agent on separate evaluation episodes and save the best checkpoint. Pair it with TensorBoard logging to track reward curves, loss, and custom metrics in real time, giving you early stopping signals and reproducible training workflows.

Stable Baselines3 supports any Gymnasium-compatible environment, including custom ones with image observations. Define your action and observation spaces, implement the step() and reset() methods, and optionally add CNN policies for pixel inputs; SB3 handles policy selection and training automatically.

Use the save() and load() methods to persist trained agents as serialized files. You can also checkpoint during training via EvalCallback, storing the best-performing model snapshot; this enables reproducible evaluation, transfer learning, and deployment workflows without retraining.

Stable Baselines3 implements PPO, SAC, TD3, DQN, A2C, HER, and other on-policy and off-policy algorithms. Each suits different problem structures: PPO and A2C for continuous control, DQN for discrete actions, SAC and TD3 for off-policy sample efficiency, and HER for sparse-reward manipulation tasks.