Fail-Safe Controller for Assisting Users of Personal Mobility Robot with BMI

Abstract

We have thus far presented a brain-machine-interface, BMI, for users of personal mobility robots. However, once the BMI predicts a wrong control command, both the user and robot face the danger of collision accidents. In this paper, therefore, we propose a fail-safe controller based on CNN (Convolutional Neural Network) for assisting users of personal mobility robots with the BMI. In addition to the control command, a depth map for the input image is simultaneously predicted by the fail-safe controller through multi-task learning. For this purpose, CAE (Convolutional Autoencoder) and DCGAN (Deep Convolutional Generative Adversarial Networks) are used instead of the CNN. In the experiments, we show that the fail-safe performance is increased by predicting the depth map for the input image. Finally, the fail-safe controller based on the DCGAN yields the best performance.