While learning techniques for control can be robust, they come at the cost of losing modularity, tunability, and understanding of the controllers. How can we gain insights about locomotion from learned policies and structure our controllers to retain these properties? How can we use model-based and model-free controllers in tandem in an optimal way?
