Training End-to-End Steering of a Self-Balancing Mobile Robot Based on RGB-D Image and Deep ConvNet

Abstract

In an attempt to build a self-balancing mobile robot, an end-to-end autonomous steering system was proposed based on RGB-D deep learning. An RGB-D camera is installed on the mobile robot to capture real-time depth images of the front environment. By annotating the depth images with the steering angles, a deep convolutional neural network was trained to provide end-to-end steering commands for direction control. The self-balancing mobile robot was built on a commercial self-balancing mobile platform; a data collection scooter was built on the same mobile platform and operated by a human rider to reflect the social-aware behavior. Two types of navigation tasks - corridor cornering and path adjustment were devised and tested. A preliminary performance study was also reported.