Learning a Structured Neural Network Policy for a Hopping Task

In this work, we attempt to learn a neural network policy for dynamic, underactuated locomotion tasks. Learning a policy for such a task is non trivial due to the dynamic, fast changing, non linear and contact rich dynamics of this task. We use existing trajectory optimization techniques to optimize a set of policies. For this, we present a method that allows to learn contact rich dynamics for underactuated systems in a sample efficient manner. Using a new kind of neural network architecture, we are able to preserve more control structure and information from the optimized policies. This way, the network output is more interpretable. We analyze the quality of the learned dynamics and the robustness as well as generalization of the learned network policy on a set of tasks for a simulated hopping task. We also inspect if the network has learned reasonable control structures from the optimized policies.