Offboard Machine Learning Through Edge Computing for Robotic Applications

The paper presents a study on offboard machine learning through edge computing for robotic applications. Edge or `fog' computing is being pursued as a supplement to cloud computing to provide the computing resources near the `edge' where it is needed most. In this paper a framework of such an edge computing system is presented for robotic applications. The system consists of a recent machine learning platform (Jetson TX2) integrated within a heterogeneous robotic environment of UAVs and mobile robots operated through robot operating system (ROS). The UAVs and the mobile robots are equipped with cameras for providing visual feedback of the surrounding to the main processor (Jetson TX2) for image processing and object classification. The UAVs and the robots have their individual local processors capable enough to handle local processing of other sensor data and providing interface for transmitting images to and receiving identification results from the offboard processor. The system has been developed around ROS which provides the communication between hardware and system drivers for UAVs and mobile robots. The effectiveness of the approach is demonstrated using a UAV and a mobile robot to send images to the offboard process to identify a moving person and receive commands to follow him. The mobile robot follows the person and the UAV action is simulated using a simulator with parameters of a physical UAV.