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

16 In conversation | Terry Gafron can build a system more like that in a house fly, he says. This neural networks approach is not new, but the memristor mechanism is. In the past, a neural network had to handle complex statistics in software with a high computational requirement. “I think the direction the technology will go will be in moving from the software algorithm to the hardware level,” Gafron says. “Instead of statistical probabilities, the neural network approach enables most of the decision- making to be done at the hardware level. The output is simply the function of the inputs, which then goes into the decision tree that is the control system. It’s doing the statistical work at the hardware level and in real time.” This is a critical step towards full autonomy, he says. Getting to the final goal is a combination of many things, a lot of which depend on the ability to do the processing in the hardware rather than the software. It can also help improve the performance of a UAV, and allow it to operate at higher speeds. “As the speed of the aircraft increases, its ability to respond in a timely manner becomes compromised – it’s moving too fast to avoid obstacles,” Gafron says. “So instead of large processing engines, the hardware has to determine what an obstacle is and whether it’s a threat. You don’t need to use large banks of sensors to identify a tree, for example.” Gafron uses the example of a UAV operating in a big city, perhaps above street level for monitoring traffic. Here, one response to it finding there’s a building in the way would be to fly over it, but flying around it would be better as you wouldn’t need to know how high the building is. In a rural environment though, that condition would probably not apply. All of this is much closer to how autonomy has to function, he says. It is also important to new ways of controlling multiple systems in a swarm, as a memristor-based neural network costs less, is lighter and uses less power. Conventional UAVs are large systems with powerful sensors such as Lidar and so on, and that requires a lot of power and adds weight and cost, but for UAVs to operate in swarms they have to be lightweight. “That’s why we think this is the only solution that can actually be implemented, and we are proposing control modules that measure 2 x 2 x 0.5 in and contain all the computational power for autonomy,” Gafron says. “We spend a lot of time designing ICs for the DoD so it’s quite possible for us to design a UAV-specific control chip whose sole purpose is autonomous control that you could implement across the board with all flight systems. People are trying to solve the autonomy problem with a 5 lb computer, but it has to use this kind of ambiguous computing.” The swarm approach raises another point. The ability of a UAV to control itself is important, but when it has limited power and endurance, how far can it fly? With a swarm technology, large tracts of land can be monitored for crop analysis using lots of smaller UAVs that are communicating with each other, sharing sensor data, and there are many ways to implement this swarm approach. By passing data back through the swarm, a larger area could be covered than by a single UAV. Bio Inspired has now started delivering chips with banks of memristors, as well as development boards for programming them, to researchers and developers around the world. “As a small company we don’t have to play the marketing game,” Gafron says. “The bigger companies have to convince their investors and their boards that they are working on the technology and have to make billions, but we don’t have to do that, we just have to make the technology and sell it.” The devices range from 20 memristors down to eight for a lower cost system, and boards to link to the chips. “What we do is enable a new kind of computing that is more like the human brain,” Gafron says. “That has always been the problem with this technology – it’s not a better mousetrap, it’s a different mousetrap. That’s why the UAV work is so important, as it can provide better computing systems that keep us safe – we can recreate learning and adaptive learning in a microchip.” Summer 2015 | Unmanned Systems Technology A resistor with memory, or memristor, was predicted in 1971 but only demonstrated in 2008. A memristor uses materials called chalcogenides, where the change of the electrical state of the material depends on the voltage applied to it. This change is cumulative, as fingers, called dendrites, in the material grow or shrink depending on the voltage. That means the change in the memristor’s resistance depends on the voltages that have been applied previously, so it can provide a key element of a neural network – the weighting. Companies such as HP and IBM are working on the technology for a range of applications, but Bio Inspired has licensed memristor technology from Boise State University to use in integrated circuits for neural networks. This adds four or five layers to a standard chip-making process, so it can be easily added to a chip design. Memristors Building memories that learn on a silicon wafer is providing new ways of controlling autonomous systems