Viability of NPU-Equipped SBCs for Real-Time Onboard Vision Autonomy on Lighter-Than-Air UAVs

UAVs face unique constraints in autonomous operation, including limited mass, power, and possible unreliable communication under degraded telemetry relay conditions. These challenges necessitate stable onboard inference, as off-board compute becomes undependable during real-time flight. In this work, we assess the viability of lightweight, Neural Processing Unit (NPU)-equipped single-board computers (SBCs) for onboard object detection and control and propose a configurable system design adaptable to a range of UAV types, including lightweight, high-speed, maneuverable platforms. We benchmark CPU-only and NPU-accelerated SBCs--specifically the Raspberry Pi Zero 2W and the Rockchip-based Radxa ZERO 3W and Khadas Edge 2, respectively--through stress tests to evaluate inference latency consistency. Following that, we present an onboard perception pipeline using a YOLOv8-nano model with minimal off-board telemetry. Our results demonstrate that modern NPUs sustain reliable real-time inference and control under the physical and environmental limitations of lighter-than-air UAVs, enabling robust autonomy without dependence on high-bandwidth ground links.