An Experimental Study of the Autonomous Helicopter Landing Problem

Image-based information is useful for a variety of autonomous vehicle applications such as obstacle avoidance, map generation, target tracking and motion estimation. Cameras are inexpensive and operate at a high temporal rate. This, coupled with advances in processing power, makes images-based techniques invaluable in autonomous systems operation. In this paper we propose and experimentally investigate a vision-based technique for autonomously landing a robotic helicopter. We model the solution to the landing problem discretely using a finite state machine, responsible for detecting the landing site, navigating toward it, and landing on it. Data from a single on-board camera are combined with attitude and position measurements from an on-board inertial navigation unit. These are the inputs to the on-board control system: a set of controllers running in parallel which are responsible for controlling the individual degrees of freedom of the helicopter. The resulting hybrid control system is simple, yet effective as shown experimentally by trials in nominal and perturbed conditions. We experimentally test our algorithm by initializing the helicopter in hover at an arbitrary location. The helicopter is required to autonomously locate a helipad (imprinted with a known visual landmark), align with it and land on it. Results from experiments show that our method is able to land the helicopter on the helipad repeatably and accurately. On an average the helicopter landed to within 35 cm position accuracy and to within 7 in orientation as measured from the center of the helipad and its principal axis respectively. In this paper we focus on experimental evidence showing the robustness of the algorithm. In particular we show that 1. the helicopter is able to visually re-acquire the helipad after losing it momentarily, and 2. the helicopter is capable of tracking a moving helipad and landing on it, once the helipad has stopped. In the tracking experiments the helipad was moved a significant distance (7 m on an average). Importantly the same algorithm is used across these conditions no specific modifications were made to handle the various cases. In the following section, we give an overview of the vision and control algorithms. Following this, a representative sample of the results obtained are shown. A detailed analysis of the assumptions made, algorithms used, and results obtained will be presented in the final paper.