**1. Introduction**

Robotics is increasingly taking on greater importance in our lives. One of the main areas where this can be perceived is Search and Rescue (SaR) tasks [1]. Robots designed for this kind of task, known as SaR robots, must operate on many occasions in unknown environments, move over unstable surfaces, and face multiple difficulties in order to carry out their mission, e.g., obtaining a map of the environment to facilitate the subsequent intervention of the rescue brigades [2]. Using a single robot under such conditions poses big difficulties: whether it moves on the surface or flies nearby areas, there are intrinsic difficulties for each type of robot. Thus, by building heterogeneous teams of robotic platforms that can jointly operate in such scenarios, it is possible to bring about great benefits, since the shortcomings of each robot can be compensated with the strengths of the other [3,4].

Indeed, while aerial robots have the unique ability to obtain top views from the terrain and move without being hampered by the elements that may be found on the ground after a collapse, their reduced flight autonomy limits their operating time to a few tens of minutes. Moreover, their load capacity is generally less than 1 kg, which limits the type of sensors or equipment that can be deployed.

On the other hand, terrestrial robots are able to overcome, in general, the requirements of energy autonomy and payload. In addition, they can act as relays for communication systems, as well as provide high computing capabilities and data storage to the system. They have, however, limited mobility, especially in cluttered environments, such as narrow bridges or inclined planes. Additionally, their ability to obtain information about their environment may also be limited by their low height above the ground level and by the very elements of the scenario.

The literature contains multiple examples of successful collaboration of ground and aerial robots to carry out different missions: exploration in wide areas with obstacles [3]; precision farming for ground moisture sampling [5]; surveillance in complex environments using route optimization strategies [6]; and supporting aerial surveys in maritime environments [7], where the maritime robot acts as a mobile landing platform of the Unmanned Aerial Vehicle (UAV) when it has to perform an emergency landing, charge its batteries, or be picked up by an operator. All these examples prove the efficiency and benefit of building mixed robotic systems comprised of a terrestrial and an aerial robot for many different and complex tasks.

This work proposes a step towards obtaining such a joint team by developing a system that enables a UAV to: (1) take off autonomously from a landing platform attached to a Unmanned Ground Vehicle (UGV); (2) detect, localize, and follow the ground robot while in the air; and (3) land autonomously on the moving platform when required.

The proposed system differs from previous works by presenting a novel height-adaptive controller for tracking and landing. In essence, the behavior of a Proportional-Integral-Derivative (PID) controller is modified according to the UAV's distance to the landing platform along the vertical axis. By doing so, the performance and robustness of the system as a whole are largely increased.

Two different approaches to track and land robustly the UAV on the moving landing platform are presented: in the first variant, the aforementioned height-adaptive controller uses the current position of the landing platform as the target; the second approach extends the height-adaptive controller with a prediction algorithm (based on a Kalman filter) that predicts the future position of the platform and feeds it to the controller. This facilitates tracking and, more importantly, the autonomous landing of the aerial robot.

Another key novelty introduced in this work is the addition of a recovery and re-localization module for both tracking and landing. This further helps to increase the robustness of the system, because the UAV can re-detect the landing platform autonomously in case the latter disappears from the field of view of the UAV's camera or if the relative error between the landing platform and the UAV in the immediate moments before landing is greater than a threshold.

Furthermore, a novel finite state machine is presented in this work, which together with our re-localization module allows for life-long operation, as we will demonstrate in Section 4.

The proposed system in its two versions has been extensively tested on a realistic three-Dimensional (3D) simulated environment (Gazebo [8]) and deployed for qualitative evaluation on real robotic platforms. Figure 1 shows the proposed aerial-ground robot fleet. We employ Robotnik's Summit XL as the UGV and the Parrot's AR Drone 2.0 as the UAV. In the simulated environment, we use the corresponding Robot Operating System (ROS) [9] packages, namely the *summit\_xl\_sim* and *tum\_simulator* packages.

**Figure 1.** Cooperation between Unmanned Aerial Vehicles (UAVs) and Unmanned Ground Vehicles (UGVs) can greatly benefit Search and Rescue (SaR) tasks where both long-term operation and a wide aerial view are required. The UAV can travel on top of the UGV (possibly recharging its battery) and, when needed, take off, inspect the area, and land again autonomously.

This paper is organized as follows. Section 2 analyzes previous work. In Section 3, an overview of our system is presented, and the robotic frameworks and platforms used are described. Section 4 presents an extensive evaluation of the system in the simulated environment and qualitative tests in the real robotic platforms. Section 5 concludes this work.
