Flying in dynamic, urban, highly-populated environments represents an open problem in robotics. State-of-the-art (SoA) autonomous Unmanned Aerial Vehicles (UAVs) employ advanced computer vision techniques based on computationally expensive algorithms, such as Simultaneous Localization and Mapping (SLAM) or Convolutional Neural Networks (CNNs) to navigate in such environments. In the Internet-of-Things (IoT) era, nano-size UAVs capable of autonomous navigation would be extremely desirable as self-aware mobile IoT nodes. However, autonomous flight is considered unaffordable in the context of nano-scale UAVs, where the ultra-constrained power envelopes of tiny rotor-crafts limit the on-board computational capabilities to low-power microcontrollers. In this work, we present the first vertically integrated system for fully autonomous deep neural network-based navigation on nano-size UAVs. Our system is based on GAP8, a novel parallel ultra-low-power computing platform, and deployed on a 27 g commercial, open-source CrazyFlie 2.0 nano-quadrotor. We discuss a methodology and software mapping tools that enable the SoA CNN presented in [1] to be fully executed on-board within a strict 12 fps real-time constraint with no compromise in terms of flight results, while all processing is done with only 94 mW on average - 1% of the power envelope of the deployed nano-aircraft.

本文介绍了一种用于小型自主飞行器（如无人机）的自主导航引擎，此引擎基于神经网络和并行低功耗计算平台GAP8，可以使得纳米无人机在严格的实时约束下完整地执行复杂的深度卷积神经网络，以实现闭环端到端可视化导航，同时系统耗能相对较小，且不会对飞行结果进行任何妥协。

基于DNN的64mW自主纳米无人机视觉导航引擎