First information about setting up this repository can be found in our readme. This wiki contains tutorials and some additional tips and tricks. Take a look at the tutorials below!
The most common use case for a researcher is probably to extend an existing algorithm with a new feature or mechanism. This is covered in the tutorial to Adding a new algorithm.
Tutorials
How to configure Pycharm for this project.
How to run a standard algorithm(SAC with Hindsight Experience Replay from stable baselines 3) and interpret the console output. This is also a good starting point for checking out other tutorials.
How to visualize and render.
How to manage hyperparameters and arguments with Hydra.
How to log data.
How to display logged data in MLFlow and Weights and Biases.
How to restore a saved policy.
How to set early stopping values.
As an example of how to create a new different algorithm from an existing one, we create a copy of CLEANSAC and start modifying it
How to perform hyperparameter optimization.
How to set up and perform a performance test.
How to set up and perform a smoke test.
How to add a new robotic environment.
How to equip an algorithm with a forward model.
Further information
We created an overview over our custom environments.
Tips for training on remote computers.
Detailed instructions on setting up the project, especially for Windows with WSL2.
How our GitLab pipeline works.
Table of contents
- Add new algorithms
- Add environment to MakeDictObs wrapper
- Add new environments
- Equip RL-algorithm with forward model
- Custom Environments
- Run an experiment
- Display logged data
- Online metric visualization
- Restore a policy
- Hyperparameter management
- Hyperparameter optimization
- Early stopping
- Performance tests
- Smoke tests
- GitLab Pipeline
- Logger
- Train on remote machine
- PyCharm
- File structure
- Detailed instructions for getting started