When it comes to quadrotors and other aerial drones, some researchers use computer vision algorithms and 3D represented environments to train autonomous drones how to navigate through that particular space. According to engineers from the University of Maryland’s Perceptions and Robotics Group, that practice is highly inefficient as they are not ‘task driven,’ and not utilized by animals and insects, who have been solving the problem of navigation without the need to build 3D models for eons.
As you might have suspected, the engineers looked to insects, bees in particular, to design a drone dubbed GapFlyt, which is capable of flying through small spaces at relatively high speeds with no training beforehand. The quadrotor only requires taking a few sensing snapshots to define the space’s opening before traveling through, even if the hole is irregular and not symmetrical.
To travel through any given opening, the drone uses a technique known as optical flow- essentially the pattern of apparent motion of objects, surfaces, and edges caused by the relative movement of the observer, in which case is the drone and the visual platform it uses to perceive those patterns.
The drone creates a 3D model of the opening using a single camera that captures images and marking the features found in each. It can then calculate the shape and depth of the opening it will travel through.
To accomplish that feat, the engineers outfitted the GapFlyt (a Parrot Bebop 2) with a single front-facing camera, a downward-facing optical flow sensor (camera/sonar), and a Nvidia TX2 module. The forward-facing camera is used to calculate the opening using patterns derived from the images, while optical flow sensor is used to hold position relative to the opening, while all data is processed using the TX2 and PRG’s TS2P algorithm.
Those looking to make their own GapFlyt drone are in luck, as the Perceptions and Robotics Group engineers have made their design open source and have uploaded all of the files and code needed to their GitHub page.