Unmanned Systems Technology 027 l Hummingbird XRP l Gimbals l UAVs insight l AUVSI report part 2 l O’Neill Power Systems NorEaster l Kratos Defense ATMA l Performance Monitoring l Kongsberg Maritime Sounder

15 Unmanned Systems Technology | August/September 2019 Kar-go has launched a production version of its autonomous delivery vehicle (writes Nick Flaherty). Kar-go uses a ‘terrain training’ approach to enable each vehicle to become an expert in its particular route so that it can then focus on any abnormalities in that route. The vehicles will operate in fleets, each one delivering within a dedicated local area, so it never needs to drive too far in an untested environment. The system has been designed by the Academy of Robotics in Wales, UK, for delivery applications. It continuously scans under other vehicles on the road to determine if there is a moving object on the other side and what direction it is moving in. This is intended to enhance safety in busy residential streets by avoiding pedestrians. The terrain training uses landmarks to get the exact position on a local route. For example, identifying a clearly visible bank 10 m away, a lamp post 10 m away and a cell tower 50 m away gives three points the Kar-go can use to triangulate its exact position, rather than relying on GNSS satellite navigation. The system uses modules rather than a single controller for functions such as lane and object detection. This modular structure means the final outputs relating to speed, steering and braking are not processed from a single module, but are the result of integrating perception and decision outcomes of multiple modules. That makes the system less sensitive to disturbances and noise. Each module is designed to operate individually as well as collaboratively, as any module can take control in case another one fails. A decision-making convolutional neural network (CNN) runs on Nvidia’s Drive graphics processing unit. However, this is not sufficient for effective operation, as it is a static, passive-vision system that perceives the world one frame at a time with no knowledge of prior frames. The CNN is also trained using a mathematical optimisation process, which is not how intelligence emerges in nature. Instead the Kar-go uses a technique called long short-term recurrent networks, which are used in other domains such as natural language processing. They use a sequence of preceding frames to predict a new frame and then compare that predicted frame with the incoming data. The controller also uses a ‘continuous time-recurrent neural network’ that also takes previous frames as the input. This algorithm changes on the basis of the content of the images it uses, evolving into a more accurate algorithm. Addressing the challenge of communicating with humans and other autonomous vehicles, the vehicle’s outer body is fitted with reflective screens and LED rings that can light up and change appearance in different scenarios. These help the Kar-go to communicate with other drivers and pedestrians during deliveries. Fleet expert in deliveries Driverless vehicles The Kar-go uses ‘terrain training’ to allow fleets of them to be operated autonomously