Improving Learning in a Mobile Robot using Adversarial Training

This paper reports research training a mobile robot to carry out a simple task. Specifically, we report on experiments in learning to strike a ball to hit a target on the ground. We trained a neural network to control a robot to carry out this task with data from a small number of trials with a physical robot. We compare the results of using this neural network with that of using a neural-network trained with this dataset plus the output of a generative adversarial network (GAN) trained on the same data. We find that the neural network that uses the GAN-generated data provides better performance. This relationship holds as we present the robot with generalized versions of this task. We also find that we can produce acceptable results with an exceptionally small initial dataset. We propose that this is a possible way to solve the "big data" problem, where training a neural network to learn physical tasks requires a large corpus of labeled trial data that can be difficult to obtain.