Low-Power Footprint Inference with a Deep Neural Network offloaded to a Service Robot through Edge Computing

Abstract

Recent advances in the field of Artificial Intelligence (AI) have enabled a vast variety of innovative digital services. Mobile smart devices usually resort to the cloud to run deep neural networks (DNN) due to insufficient computational power or severe power constraints that precludes the use of consumer-grade on-board processors and power-hungry Graphics Processing Units (GPU). However, the use of cloud computing in service robot deployments has shortcomings related with latency, privacy, security and reliability, which often makes it inconvenient or even impractical. A possible solution is the use of specialized edge computing devices with a trade-off between onboard robot computing resources and power footprint. This approach is exploited in this paper for a service robot programmed in ROS, equipped with a camera for image perception, a 2D LiDAR for autonomous navigation, and a system on module Nvidia Jetson AGX Xavier. The viability of running DNN aboard this robot to perform image classification with low-power footprint in a Covid-19 use case scenario is demonstrated.