Page - 7 - in Proceedings of the OAGM&ARW Joint Workshop - Vision, Automation and Robotics

Package Delivery Experiments with a Camera Drone Jesús Pestana1 and Michael Maurer1 and Daniel Muschick2 and Devesh Adlakha1 and Horst Bischof1 and Friedrich Fraundorfer1 Abstract—The undergoing efforts for the integration of robotics into logistics systems is affecting the production work- flow at all stages, from the transportation and the handling of parts insidestorageandproductionfacilities to thefinalproduct distribution. In this paper we address the problem of delivering a package by means of a multirotor drone. We describe a fully autonomous package delivery flight demonstration prepared in collaboration with an industrial partner. All computations are performed in real-time on-board the drone. A gimbal camera is utilized to realize the vision-based localization, by means of fiducial markers, of the delivery position and the landing platform on a pickup truck. The demonstration consists of the fully autonomous execution of the following tasks: the drone takes-off from the truck, looks for the delivery position, proceeds to land and drop the package, flies back to the distribution truck and follows it, and the flight is finished by performing the landing on the static vehicle. The experiments focus on the performance of the vision-based truck following. I. INTRODUCTION In this paper we present a fully autonomous drone that using only on-board processing is able to perform coarse navigationusingGPS, vision-basedprecisevehicle following and landing on static platforms (see Figs. 1 & 3). We used our system to perform a fully autonomous package deliv- ery flight demonstration in collaboration with an industrial partner. The main technical challenges related to this work are the navigation control, the real-time vision-based pose estimation of the vehicle and the landing positions and their integration with the navigation control. In order to obtain the required localization precision for the vehicle following and the landing tasks we use visual fiducial markers. Drones are a hot topic and an ongoing research area. These aerial platforms are suitable for being integrated in logistics systems, for instance, for the transportation of goods. Package delivery by means of an autonomous drone can significantly reduce the costs of distribution. A succinct feasibility analysis by D’Andrea [4] estimated its operating cost at 10 cents for a 2 kg payload and a 10 km range. The main challenges faced by real-world drone package delivery are highlighted by the following selection of recent research works: an obstacle mapping method that encodes at cell-level the value of occupancy and its variance [1], testing modern deep-learning based object detection algorithms on- board drones [6], trajectory planning intended for navigation in cluttered environments [3] and landing on vehicles that are moving in straight roads at speeds of up to 40 km/h [2]. 1Institute for Computer Graphics and Vision, ICG - TU Graz {pestana,maurer,bischof,fraundorfer}@icg.tugraz.at 2Institute of Automation and Control, Graz University of Technology daniel.muschick@bioenergy2020.eu Fig. 1. Illustration of autonomous vision-based controlled drone landing on a marked delivery position in order to deliver a package. DJI M100 Autopilot Board GPS Module NVIDIA Jetson TK1 - Quad-core ARM Processor - NVIDIA Kepler GPU Gimbal and Camera DJI Zemuse X3 Fig. 2. DJI M100 quadrotor equipped with a Nvidia Jetson TK1 on-board computer (DJI Manifold), autopilot, GPS module, DJI Zenmuse X3 gimbal camera (1280×720 px) and an electro-magnet to carry a package of 100 g. II. SYSTEM OVERVIEW a) Hardware Setup Our drone is equipped as shown in Fig. 2. For the exper- imental tests, the E-Mobility electric powered pickup truck “ELI” from SFL Technologies [7] was used, see Fig. 3. The delivery position and the landing platform are tagged using a 39×39 cm 36h11-family Apriltag fiducial marker [5]. Using this approach the relative pose of the gimbal camera with respect to the landing-platform at a distance of 3.5 m can be estimated with an accuracy of around 3 cm. b) Software Setup The inter-module communication is achieved by means of the Robot Operating System (ROS). Since our experimental results focuson thecar followingperformance,only themain modules related to this task are explained, which are: the gimbal camera landing-platform tracking, the vehicle speed estimation and the control algorithm. 7