Unmanned Systems Technology 003 | UAV Solutions Talon 120 | Cable harnesses | Austro Engine AE50R and AE300 | Autonomous mining | AUVSI 2015 show report | Transponders | Space systems

12 Platform one Summer 2015 | Unmanned Systems Technology Neural networks have been around for decades, but new materials and processor designs are opening up the technology for applications such as unmanned air and ground systems. A small US company has now launched commercial devices called memristors, which can be used to provide a key element of a neural network to control UAVs. As highlighted in the ‘In Conversation’ article in this issue, Bio Inspired Technologies, of Boise, Idaho, is selling memristors on a die for researchers, but it is also launching a USB board that will allow more detailed development of control systems around the technology. Bio Inspired is not the only company looking at memristors. HP has been working on the technology since 2008 and is planning to launch an operating system based on the devices in 2015. Meanwhile, researchers in the US have devised a method of producing a three-terminal memristor that could act like a transistor. The research, led by Professor Mark Hersam of Northwestern University’s McCormick School of Engineering, uses a single layer of molybdenum disulphide (MoS 2 ), a two-dimensional nanomaterial semiconductor with a granular structure. When a large electric field is applied to the material, the grain boundary physically moves, causing a change in resistance. By using MoS 2 with this grain boundary effect instead of the more usual metal-oxide, metal memristor structure, the team has demonstrated a new three-terminal memristive device that can be tuned using a gate electrode. “With a memristor that can be tuned with a third electrode, we have the possibility of realising brain-like computing,” said Prof Hersam. “We are now actively exploring this possibility in the laboratory.” In an another development, Synopsys has devised a more traditional way of implementing low-power neural networks for image processing in embedded systems. Its EV52 and EV54 chips are fully programmable and configurable HD vision processor cores that would be integrated into an on-chip system. They implement a convolutional neural network (CNN) that can operate at more than 1000 GOPS/W, allowing fast and accurate detection of objects such as people that may be encountered by an autonomous system. “In embedded systems, power consumption becomes a big deal, and a graphics processor with lots of cores has a problem with power, so dedicated vision processors have been starting to show up over the past few years,” said Mike Thompson, senior vice-president of marketing at Synopsys. “While CNNs have been around for a long time they were modified in about 2005 to improve their efficiency and accuracy, to the point where Microsoft and Google announced that the networks have over 95% accuracy for image recognition, which is better than people can achieve,” he said. The processors use four RISC CPU cores with up to eight separate processing elements which are user-configurable and which can be programmed to run CNN algorithms. The training is done offline, but the weightings, or executables, are programmed into the object detection engine, allowing the engine to recognise an HD camera image. The processing elements are customised for handling CNN weightings; using configurable technology acquired from ARC International they are still programmable with instructions and data memory. This has allowed Synopsys to port popular image recognition software such as OpenCV to the architecture to make it easier for developers to use the chip. “These are the first of a series of products, and they will have some use in basic ADAS-assisted driving applications, but vision requirements such as lane detection requires vector DSP processing, and that’s a slightly different engine,” said Thompson. Resistors that can remember Networking The memristor development board from Bio Inspired Technologies