Learning Optimal Robot Ball Catching Trajectories Directly from the Model-based Trajectory Loss

This paper is concerned with learning to compute optimal robot trajectories for a given parametrized task. We propose to train a neural network directly with the model-based loss function that defines the optimization goal for the trajectories. This is opposed to computing optimal trajectories and learning from that data and opposed to using reinforcement learning. As the resulting optimization problem is very ill-conditioned, we propose a preconditioner based on the inverse Hessian of the part of the loss related to the robot dynamics. We also propose how to integrate this into a commonly used dataflow-based auto-differentiation framework (TensorFlow). Thus it keeps the framework's generality regarding the definition of losses, layers, and dataflow. We show a simulation case study of a robot arm catching a flying ball and keeping it in the torus shaped bat. The method can also optimize "voluntary task parameters", here the starting configuration of the robot.