AI GitHub
Stable Baselines3 for standard reinforcement-learning algorithms
Train and evaluate agents using established PyTorch implementations.
Key Takeaways
- The library implements algorithms around Gymnasium environments and adds utilities for monitoring and vectorizing runs.
- Select an algorithm that supports the environment's action space.
- Train a simple baseline over several seeds and evaluate it in a separate environment.
- Learning curves can vary sharply across seeds and reward definitions.
How it works
The library implements algorithms around Gymnasium environments and adds utilities for monitoring and vectorizing runs. Algorithm choice depends on the action space and sample budget.
What you need
Select an algorithm that supports the environment's action space. Use a separate evaluation configuration and record any observation or reward transforms in both paths.
What to try first
Train a simple baseline over several seeds and evaluate it in a separate environment. Save the wrapper configuration with the policy.
Run an evaluation without updating observation-normalization statistics. If the statistics keep changing, comparisons across saved policies can be misleading.
Where to look for failures
Learning curves can vary sharply across seeds and reward definitions. Keep normalization and evaluation settings fixed when comparing agents.
Related Franklin pages
Separate the physics simulator, environment interface and learning algorithm. Test resets and termination before increasing the number of workers or training steps.
- Configure simulated robotics tasks for reinforcement or imitation learning. Isaac Lab for GPU-accelerated robot learning
- Model articulated systems, contacts and actuators in a physics simulation. MuJoCo for simulating robot dynamics
Franklin’s robotics coverage gives examples of where simulation, training and hardware meet. Franklin’s robotics coverage
Project documentation
Try it for your workflow
Explore the project’s documentation for setup and usage, or save this page to a collection for later.
Read the project documentation ↗