USE Network launch I UAV Works VALAQ l Cable harnesses l USVs insight l Xponential 2020 update l MARIN AUV l Suter Industries TOA 288 l Vitirover l AI systems l Vtrus ABI

82 A rtificial intelligence (AI) for unmanned systems is a combination of several types of technology, all with different advantages but with similar constraints on size, weight and power consumption. Rule-based expert systems and path prediction algorithms have been widely used for many years to provide AI, but now various kinds of neural networks are being added to the mix of design options for autonomous systems. The key capability of a neural network is pattern recognition. This can be for image recognition – identifying road signs, pedestrians or road vehicles, aircraft or ships – or collision avoidance, and is one of the major drivers of the technology these days. Pattern recognition can also be used to monitor the condition of a battery pack to predict its performance, and for the control system by collecting millions of kilometres of real and simulated travel to create a detailed model for controlling a vehicle. Neural networks are also being developed to identify the arm movements of humans. This would allow a marshal at an airport or aircraft carrier apron for example to control the movements of a UAV in the same way as a piloted aircraft. Training a neural network is carried out on large, power-hungry processors in a rack or data centre to iteratively build and test models in different types of neural network frameworks. The resulting models can then be run on fixed networks called inference engines, often using similar processors to the ones in the data centre. The current convolutional neural network (CNN) and deep neural network (DNN) implementations use massive arrays of multiply accumulate (MAC) units that are commonly found in digital signal processors used for video processing, graphics processing units (GPUs) and field programmable gate arrays (FPGAs). One thing that is different is that these arrays of MAC blocks require more complex memory structures, which can consume a lot of power. There is a growing range of ways to provide AI for unmanned systems, writes Nick Flaherty , who explains the key developments Intelligence agencies June/July 2020 | Unmanned Systems Technology Event-driven spiking neural networks can dramatically reduce the power needed for pattern recognition in radar systems in UAVs (Courtesy of imec)