UC San Diego

Unmanned Aerial Vehicles (UAVs) are widely used in many civilian and military applications such as package delivery, precision agriculture, mobile edge computing, and reconnaissance. In these applications, UAVs often need to perform computationally expensive tasks such as path planning, object detection, or mobile computing services. However, due to the small payload, the amount of computing resources individual UAVs can carry is limited. Although significant advances have been made in improving UAV technologies from aspects such as mechanics, control, communication, and networking, enhancing the onboard computing capacity of UAVs hasn't gained much attention as above mentioned aspects. This dissertation aims to fill this research gap by exploring Networked Airborne Computing (NAC), a new computing paradigm that aims to achieve high-performance airborne computing via inter-vehicle resource sharing using direct flight-to-flight communication links. We first investigate how to enhance the onboard computing capacity of individual UAVs in Chapter 2 by designing the onboard hardware and software. As the computing capability of individual UAVs is still limited due to small payload, Chapters 3 and 4 further explore how to leverage resources from neighboring UAVs to enhance a UAV's airborne computing capacity by using distributed computing techniques. Particularly, Chapter 3 investigates the static scenario where UAVs hover in the air while conducting computations. To optimize airborne computing performance, a coded distributed computing framework is introduced. Chapter 4 extends the analysis to dynamic scenarios where UAVs are in motion during computation, and a mobility-aware coded distributed computing framework is proposed to address these scenarios. The computation problem considered in both chapters is fundamental matrix multiplication problem, which serves as building blocks for many other advanced computation problems. In Chapter 5, we shift our attention to a more complicated problem, multi-agent reinforcement learning (MARL), and investigate how to reliably and efficiently train MARL over NAC systems. Finally, Chapter 6 designs and implements a realistic simulator and hardware testbed, and conducts experiments to evaluate the performance of NAC in two computation applications including distributed matrix multiplication and distributed gradient descent. Our experiments offer valuable insights into NAC and provide guidance for future advancements.