Issue 40 Unmanned Systems Technology October/November 2021 ANYbotics ANYmal D l AI systems focus l Aquatic Drones Phoenix 5 l Space vehicles insight l Sky Eye Rapier X-25 l FlyingBasket FB3 l GCS focus l AUVSI Xponential 2021

36 M achine learning is increasingly used in different areas of unmanned systems, with a range of implementations depending on the use case. Massive machine learning neural network frameworks have been developed for image classification in large central processors, which currently consist of thousands of processing units in a large array and many megabytes of on-chip memory. The frameworks are trained offline with tagged images from a range of sources. The sources range from a ‘golden’ database of known examples, using data from sensors of vehicles on the road such as a camera, radar or Lidar, or created by simulation models to test out all the different use cases as synthetic data (see Simulation focus, UST 39, August/September 2021 for more details about the technology behind this). Once trained, the frameworks are then run as inference engines on a local processor on an unmanned system to compare the incoming data with the trained data. When used as the central control in a vehicle such as a driverless car, the neural network classifies objects in the incoming data. In an urban setting, these objects can be road markings, pedestrians, other vehicles, cyclists, street signs, potholes in the road or any other everyday object that could have an influence on the progress of the vehicle. In an off-road setting these can be very different objects, while in a maritime setting or in the air, it is more about identifying potential obstructions as early as possible. The challenge has been to increase the range of objects that can be reliably Learning curve Nick Flaherty reports on how machine learning is being implemented in an ever-growing range of applications in unmanned systems October/November 2021 | Unmanned Systems Technology Machine learning with data and control of swarms of spacecraft are being explored for mining asteroids (Courtesy of NASA/University of Arizona)

RkJQdWJsaXNoZXIy MjI2Mzk4