Autonomous Mobile Robot as a Cyber-Physical System Featuring Networked Deep Learning and Control

In this article, we proposed an autonomous mobile robot (AMR) capable of conducting multiple deep learning inferences by a computing network as a Cyber-Physical System (CPS). In view of industry 4.0, a five-layered CPS architecture including Component, Intelligence, Cyber, Configuration, and Deployment was devised. Mechatronic control of the two-wheeled self-balancing AMR presents the first challenge to embedded computing, whereas multiple deep learning modules for visual navigation and localization further escalate the stake. To meet the challenge of real-time and affordable computing, we constructed an embedded cluster based on Raspberry Pi4s and Intel NCS2 for simultaneous intelligent inference and data exchange. We tested the performances of three networking methods, Network File System (NFS), Socket Server with Multiple Clients (SSMC), and ROS TO PIC (ROS-T), and reported the comparison. We also conducted a series of field tests on autonomous corridor navigation, localization, object delivery, and object pick-and-place to verify the performance of the self-balancing AMR. The results showed high success rates with sufficient spatial precision, stability, and robust loop closure. The web-based console realizes the functions of the digital twin and provides connectivity to higher-level management systems.