Imitation Learning based on Data Augmentation for Robotic Reaching

In this paper, we present an end-to-end motion planner for a reaching task focusing on factory automation with material-handling robots. For a robot equipped with a camera sensor, convolutional autoencoder, CAE, and recurrent neural network, RNN, are used in the reaching motion planner. Imitation learning is adopted for network training of the motion planner. The robot is instructed to reach the hand toward an object beforehand. The images and instructed motions are registered by the robot and used as a training data set in the learning phase. In a framework of imitation learning, however, it is difficult for the robot to plan other motions different from the instruction. Thus the robot might not be able to reach the hand toward the object if the position is changed. For this problem, data augmentation with noise injection is applied to the training data set. In the experiment, we show that the motion planner enables the robot to reach the hand toward the object placed at not only the instructed known positions, but also unknown positions, through imitation learning based on data augmentation.