If you own a consumer drone, chances are you crashed the thing a lot when learning to fly it. Well, the same thing happens when the pros are teaching drones to fly autonomously. Those wipe-outs could become a lot less common, however, thanks to the Flight Goggles virtual reality system.
Ordinarily, when scientists are developing navigation algorithms that allow drones to avoid obstacles on their own, those drones have to fly through a space filled with such obstacles. And because the algorithms are works-in-progress, the drones don't avoid all the obstacles – at least not at first – resulting in plenty of crashes. Those mishaps can be costly and time-consuming, as the aircraft have to be repaired or replaced.
That's where Flight Goggles comes in.
Designed by an MIT team led by associate professor Sertac Karaman, it has drones flying through a large empty real-life room, tracked by motion-capture cameras that note the orientation of the drone in three-dimensional space. At the same time, a computer running an image-rendering system transmits interactive first-person-view video of an obstacle-filled environment to the drone, which it "sees" in place of the feed from its onboard camera. Each virtual obstacle is assigned a physical location in the actual room.
As a result, the aircraft attempts to avoid the computer-generated obstacles, with the motion-capture cameras determining if it succeeds in doing so. If it doesn't succeed, however, it doesn't actually crash into anything in the real world.
"The drone will be flying in an empty room, but will be 'hallucinating' a completely different environment, and will learn in that environment," says Karaman.
In a test of the system, Flight Goggles was used to train a drone to fly through a virtual window about twice as large as the aircraft itself. Throughout the course of these training sessions, its navigation algorithm self-tuned itself via trial and error. The drone subsequently utilized that algorithm to fly through an actual window the same size and location as the virtual one, guided by its onboard camera. It made it through the window 119 times, only crashing or requiring assistance an additional six times, which was part of the intended continued learning process.
Karaman and his team will be presenting their research next week in Brisbane, Australia, at the IEEE International Conference on Robotics and Automation.
Check out the system in action:
Source: MIT
