Swiss engineers have developed a control algorithm for racing drones, allowing them to autonomously fly in a changing environment, for example, along a track with moving control points. One of the main achievements is that this algorithm does not need an exact path model and can work on a board installed on the very drones, the authors of the work, published on arXiv.org, tell.
Drones used in racing competitions are piloted by radio, which can be unstable due to interference. Some engineers are developing multicopters capable of moving through the obstacle course themselves, but so far they are not perfect. Many such developments can move only along a static route and compare their current position with its high-precision model. But even with such a scheme, an autonomous flight is not easy - pictures from a moving dron may be "blurred", which makes it difficult to locate. In addition, neural network algorithms used to process images from a camera are usually too resource-intensive, so that the calculation is carried out directly on the aircraft.
Developers from the University of Zurich and Intel under the leadership of Davide Scaramuzza (Davide Scaramuzza) have developed an algorithm that allows them to autonomously navigate the race track with moving control points, using only their own resources for calculations. The algorithm consists of two main parts - a neural network that recognizes the scene and sets a goal for further movement, as well as a low-level alogorithm that plans the route and directly controls the drone.
For the first component, the authors used the modified eight-layer residual neural network ResNet-8, which they used in a previous paper devoted to training drones in the city on the basis of records from cars and bicycles. The neural network receives a color frame from the camera at a resolution of 300 by 200 pixels, and then indicates the target and the desired speed on it. These data are transmitted to the route scheduling algorithm, which is responsible for the movement of the drone directly.

The developers trained the neural network both on the data from the simulated route, on which the algorithm should adhere to the intended optimal trajectory, and from the real track. In this case, the drone flew on manual control and recorded the optimal route. As a hardware platform, the engineers used a homemade quadrocopter equipped with an Intel UpBoard for computation board and a Qualcomm Snapdragon Flight Kit for visual-interdisciplinary odometry.
The developers demonstrated the process of drone training and testing on video: