The Swiss startup Daedalean — which is developing an artificial intelligence (AI)-based autopilot for future autonomous aircraft — has selected Unigine’s 3D engine to help train the neural networks at the core of its computer vision system.
As previously reported by eVTOL.com, Daedalean is betting that computer vision enabled by convolutional neural networks will enable both near-term pilot assistance systems and, in the long term, autonomous aircraft that are safer and more reliable than human pilots. However, to learn how to react appropriately to the enormous variety of situations they might encounter in flight, these algorithms require millions of hours of training — more than can feasibly be achieved in the air.
To that end, Daedalean is largely relying on simulation to teach its neural networks not only how to recognize in-flight obstacles such as birds and wires, but also how to analyze the underlying landscape for suitable emergency landing zones: distinguishing, for example, between swamps and fields, or parking lots and roofs. According to Daedalean CEO and founder Luuk van Dijk, this type of testing “is of vital importance for obtaining full autonomy — a key enabler for scalable urban air mobility.”
Daedalean chose Unigine as the base for creating its virtual simulation environment after an extensive trial phase. Key factors in the selection were the Unigine platform’s “extremely precise coordinate system, uncompromising level of detail even at the horizon, lightning fast object loading, and perfectly realistic atmospheric model,” Daedalean said.
Unigine has extensive experience in aerospace, having worked with organizations including RSC Energia, the German Aerospace Center, and Korean Aerospace Industries, among others. It offers a version of its software development kit tailored specifically for applications in aerospace and ground transport simulation that uses a realistic ellipsoid representation of Earth, allowing aerospace users to correctly simulate “great circle” trajectories. Unigine also provides advanced projection tools, support for a variety of virtual reality headsets, and native support of thermal and night vision.
“Usually our engine is used to build simulators for real human pilots, but it works just as well for training artificial neural networks,” stated Unigine CEO Denis Shergin in a press release. “We at Unigine are happy knowing that our work could be instrumental in bringing to us the future of urban transportation, and will support Daedalean in anything they might need to fine-tune their simulation software.”