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

41 AI systems | Focus between what needs to be calculated on the UAV and what can be sent to the central engine. That then depends on the time available to make a decision, whether there are 5 seconds available or less. This approach can also integrate manned aircraft and traditional air traffic control systems that also operate in the same time frames, rather than milliseconds. Security and reliability The same pattern recognition approach used for condition monitoring can also be used for security applications. A security AI framework can monitor the flow of data through the processor using a baseline of ‘normal operation’. The same AI capabilities used in the central neural network – looking for hazards, identifying potential problems and keeping the platform stable – can be used for security and reliability. A neural network can monitor the operation of the cores, the interconnect and the I/O lines to validate the system is safe while running, and identify and classify any potential changes. It can also monitor the system for any potential failures, and adapt to compensate. For example, if a software module is compromised, an alternative module can be established on another processor core. Similarly, if a core is failing or compromised, a module can be established on another core. This leads to a more structured model of the unmanned platform, as these modules will almost certainly have real- time operation with specific deadlines. The formal model has to identify how another module can be used and still meet the worst-case timing requirements. This is the same as building neural nets that process data in a structured way. This has implications for the software architecture. The portability of AI frameworks from one hardware resource to another means more use of self-contained modules of software called containers and sophisticated orchestration software. This is now being explored in the early stages for automotive designs through the Scalable Open Architecture for Embedded Edge (SOAFEE) to standardise and prototype such systems. A 32-core system has been built for lab-based development that is capable of running autonomous workloads. It allows developers to use AI accelerator hardware to support the ARM- based CPUs. For vehicle prototyping and testing, an 80-core system has higher CPU performance, extra I/O capabilities and includes a safety processor to enable in- vehicle execution using real sensors. However, there is also the challenge of moving AI software modules from Unmanned Systems Technology | October/November 2021 Using machine learning improves the quality of wireless signals to give longer reach for UAV applications such as planting trees (Courtesy of Doodle Labs)