Modular Neural Network Policies for Learning In-Flight Object Catching with a Robot Hand-Arm System.