Deep Learning Inference at the Edge for Mobile and Aerial Robotics

Abstract

Deep learning inference is an established element for autonomous robots especially in the domain of safety, security, and rescue applications. Autonomous functions based on deep learning inference are considered the spearhead of such robots paving the way for increased demand in onboard computational resources and performance accuracy. Edge computing is steadily improved in terms of computational power but at the same time artificial neural networks are becoming even more deep and complex. To this end, the right selection between the on-board hardware platform and an efficient deep neural network is considered a challenging tradeoff issue in the autonomous system design. In this paper, we investigate the current landscape in deep learning inference at the edge by evaluating the requirements, challenges and available solutions for service-oriented architectures in the safety, security, and rescue domain. Current research directions and best optimization practices are discussed, enriched with computational and accuracy comparisons providing necessary insights for optimal system design.