A performance characterization of AI algorithms on energy-efficient hardware with applications to robust autonomous landing

Artificial Intelligence (AI) is penetrating industries on many levels, and a European wide project consisting of the main industrial leaders are working on new chip architectures and algorithms to enable very low power frameworks. Boeing Aerospace Spain (BAS) has created a novel use case in the Transport and Smart Mobility sector together with other national partners dedicated to Robust Autonomous Landing (RAL), in which AI algorithms considered vital for Unmanned Aerial Vehicle (UAV) are deployed to energy-efficient hardware. We have implemented diverse Deep Neural Network (DNN) to address RAL by means of tasks related to runway and obstacles detection and communication parameter processing. Once implemented, the models have been deployed on Field-programmable gate array (FPGA)-based accelerators to assess the reduction in power consumption, compared to traditional Graphics Processing Unit (GPU) hardware. The reductions in power consumption presented in this work outline a great potential to explore these technologies as enablers for real-time autonomy applications, generally in problems where the use of DNNs in low power scenarios is required.