An MIT drone has been filmed as it skillfully manages to avoid obstacles, at speeds reaching 30 miles per hour.
The unmanned aerial vehicle (UAV) was invented by PhD student Andrew Barry, at the Computer Science and Artificial Intelligence Lab (CSAIL), of the Massachusetts Institute of Technology.
The invention incorporates an innovative system of detecting obstacles such as trees or buildings. This allows the flying vehicle to avoid incoming objects even as it travels at high velocities.
“Everyone is building drones these days, but nobody knows how to get them to stop running into things,” declared Barry when asked about the driving force behind his work.
As he explained, lidars which assess distance using lasers are too impractical and heavy to be used on UAVs, and this problem persists for other similar sensors as well. Moreover, it’s too time-consuming and even impractical to map every area through which the drones will have to fly.
Regular programs make use of photographs taken by UAV cameras, in order to analyze surroundings at several distances (one meter, two meters, and so forth). This requires intensive computing, and significantly limits the speed of the robotic vehicles.
Therefore, the student focused on developing superior software, to equip the drone with better precision and speed, and allow it to navigate intelligently even in unfamiliar settings. That’s how he created a stereo-vision algorithm, through which drones can build a map of their surroundings in real-time, while steering clear of obstacles.
A smaller array of readings has to be processed, referring to distances of 10 meters, and as a result the UAVs can zip through the air at much higher velocities.
The open-source program now available on Github is 20 times faster than previously invented software. Camera feeds are broadcast at 120 frames per second, and depth data is analyzed at speeds reaching 8.3 milliseconds per frame.
The effectiveness of the algorithm was tested on a custom-made drone, whose components cost around $1,700. Barry equipped this UAV with two processors (similar to those on a Samsung Galaxy S3) and a camera on each wing, for stereoscopic vision.
The final product weighed as little as 664 grams and measured 34-inch in wingspan. In a demonstration, the PhD student proved how agile and efficient the device can be as it flies and avoids a multitude of trees at a farm in western Massachusetts.
However, Barry still believes there is room for improvement, and is now trying to make further adjustments so that drones can be just as reliable in dense forests and other crowded spaces.
Moore’s Law specifies that processing power tends to double every 2 years. Therefore, by taking advantage of hardware innovations, it will be possible to assess multiple depths simultaneously, and as a result algorithms will have a much lower standard deviation.
Drone navigation accuracy will also be enhanced by incorporating automatic braking systems, which are already being developed for regular cars.
Image Source: CSAIL