Multirotor, a UAS manufacturer based just outside of Berlin, has developed a new flight controller built around an architecture meant to better embed advanced algorithms and machine learning-based functions than conventional avionics, writes Rory Jackson.

The CoPilot flight controller physically packages a real-time core for flight control firmware together with an application processor as a single, physical unit. This enables high-level, AI-based behaviours to integrate seamlessly – and with no significant bandwidth limitations – with the main flight-critical autonomy stack.

“By the two memory-sharing together, the lower and higher autonomous functions can communicate instantaneously, rather than being constrained by the lengths or the bandwidth caps of a physical, conventional cable running between an isolated real-time flight controller and its edge AI computer,” said Marc Gluba, CTO of Multirotor.

Accordingly, data from cameras, Lidar and other sensors can interface directly into the flight stack and AI analytics at once, via USB or Ethernet, as can actuators via PWM, DSHOT, CAN or UART. Intelligent behaviours such as pursuit or avoidance of certain recognised objects, or agile high-speed flight between objects like trees near ground level, are hence more achievable than with conventional approaches.

The heart of the CoPilot product is an NXP i.MX8M Plus, an embedded system with an ARM Cortex-M7 for real-time flight control, four Cortex-A53 application processor cores and an integrated neural processing unit for greatly accelerating onboard AI tasks such as object detection, sensor fusion and obstacle avoidance.

“It’s one of the few embedded systems that seamlessly packages such calculation power together with a real-time processor core,” Gluba added.

The standard CoPilot measures 48 x 48 mm, and weighs 12.2 g without a heat sink. An alternate version is also available that integrates larger (and hence, for some, more convenient) connectors, as well as a wider range of interfaces, such as two MIPI CSI-2 4-lane camera interfaces.