The tutorials are divided into four sections: Foot tricks, Head tricks, Penalty taking and stopping, and Robot tricks. However, only the first tutorial is currently available, with the others marked as "coming soon". The article also notes that most of the results can be achieved in a few hours of GPU training, sometimes less, and for free on Kaggle. The notebooks used in the tutorials were originally adapted from Mujoco's MJX tutorial.
Key takeaways:
- The tutorials teach how to train a virtual humanoid/robot to perform football tricks using Reinforcement Learning.
- They provide guidance on tuning training hyperparameters for better model performance and optimizing reward functions for faster learning.
- Most of the results can be achieved in a few hours of GPU training, sometimes less, and for free on Kaggle (30h per week usage limit).
- The tutorials include foot tricks, head tricks, penalty taking and stopping, and robot tricks, some of which are coming soon.