| dc.description.abstract | The benefits of autonomous mobile platforms, such as Unmanned Aerial Vehicles (UAVs) equipped with onboard cameras, are enhanced by compact edge accelerators that are co-located, such as the NVIDIA Jetson with 100s of CUDA cores. They enable rapid inference of Deep Neural Network (DNN) models and computer vision algorithms to support real-time analytics workflows for diverse domains, ranging from smart crop monitoring and wildfire management to assisting Visually Impaired People (VIPs).
However, programming and orchestrating such UAVs/drones, either individually or as part of a fleet, and using edge computing devices for efficient, resilient and responsive operations poses challenges. The limited compute capacity of edge devices needs to be intelligently complemented by the cloud and fog computing continuum. This requires intelligent offloading strategies for compute-intensive analytics initiated by drones. At the same time, routing fleets of drones to accomplish complex tasks while leveraging heterogeneous compute requires us to optimize for and adapt to network variability, edge failures, latency and energy constraints, and monetary costs. Further, these need to be intuitively programmable to design practical applications across distributed autonomous platforms and edge resources. We address challenges of domain-aware scheduling and programmability of autonomous systems and edge platforms in this dissertation.
First, we design a cost and time-efficient task scheduling strategy, GEMS, that performs real-time decisions for DNN inference tasks generated by drones, to execute them either on edge accelerators co-located with the drone or on remote cloud resources. The goals are to maximize the Quality of Service (QoS) and Quality of Experience (QoE) for a VIP assistive application within the deadline constraints of each task. We consider the task deadline, cloud and edge pricing, and dynamic network variability to leverage strategies such as task dropping, work stealing and migration, and dynamic adaptation to cloud variability. Our realistic experiments using up to 84 emulated drones show up to 2.7× higher QoS utility, up to 75% higher QoE utility, within ±1m/s3 Jerk, up to 42% lower yaw error, and a task completion rate of up to 88% compared to state-of-the-art baselines for diverse computer vision workloads.
Next, we study co-scheduling of DNN inferencing and drone routes for a drone fleet, with inferencing performed on onboard edge devices, stationary or mobile fog devices, and public clouds. This targets collaborative applications such as smart agriculture. The drones need to visit a set of waypoints to collect data and perform analytics, with the choice of computing onboard the edge platform, on stationary fog compute at cell towers, and on mobile fog devices on public buses with specified routes. We define this as a Mission Scheduling Problem, which is NP-hard, and design the MARC scheduler as a divide and assign heuristic to solve this optimization problem. We conduct simulation-based evaluation of MARC with fleets of up to 50 drones over spatial regions based on 8 global cities and for DNN models relevant to agriculture. MARC achieves a 100% task completion rate and up to 31% higher average utility than contemporary baselines, and is within 75% of the optimal solution solved using MILP, which is tractable only for small inputs.
Further, we explore resilient scheduling of drone fleets to ensure continuity of service despite drone and edge failures in the context of wildfire response, where a heterogeneous UAVfleet helps detect stranded individuals and generate evacuation routes to safety. We develop the AeroResQ platform with algorithms that dynamically adapt to edge failures. We develop load balancing strategies to handle the active requests sent to failed drones and re-assignment of spatial regions across the available drones to ensure robust and uninterrupted orchestration. Our evaluations, conducted using an emulated environment based on recent Southern California wildfire data, demonstrate the robustness of our platform under failure scenarios and fleet configurations. The system achieves real-time performance with ≤ 1s end-to-end latency per evacuation request, much below the 2s request interval, while maintaining over 98% successful task reassignment and completion.
Finally, we design the Ocularone platform as an integrated Drones-as-a-Service (DaaS) programming and runtime framework. We design preliminary programming abstraction to enable rapid development analytics-driven UAV applications, with sensing, navigation and execution across the edge-cloud continuum using scheduling strategies we have developed. We develop the Ocularone-Bench benchmark suite to address the lack of curated datasets for uniquely identifying individuals in crowded environments and the need for benchmarking DNN inference times on resource-constrained edge devices. We also design NeoARCADE (Neo), a robust calibration technique using depth maps to estimate absolute distances to obstacles in a campus environment. As the VIP is moving, Neo uses a dynamic update method to detect inaccuracies and recalibrate the initial scale and shift parameters to adapt to the changing scenarios. These are validated in an integrated manner for a VIP assistive application to help navigate visually impaired individuals within a campus using buddy drones.
Overall, this dissertation explores diverse applications of autonomous mobile platforms operating across the edge–cloud continuum, and develops orchestration methods addressing specific research challenges motivated by these scenarios to make novel and impactful contributions. | en_US |
