Slashdot Mirror
MIT Drone Autonomously Avoids Obstacles At 30 MPH (roboticstrends.com)
An anonymous reader writes: Traditional obstacle-avoidance software uses images from each camera, and search through the depth-field at multiple distances to determine if an object is in the drone's path. Such approaches are computationally intensive, meaning the drone can't fly faster than 6 miles per hour without specialized processors. PhD student at MIT’s Computer Science and Artificial Intelligence Lab, Andrew Barry realized that at the speeds his drone could travel, the world simply does not change much between frames. Because of that, he could get away with computing just a small subset of measurements — distances of 10 meters away. "As you fly, you push that 10-meter horizon forward, and, as long as your first 10 meters are clear, you can build a full map of the world around you," Barry says.
The video shows the detected obstacles, and the drone by far fails to detect all relevant obstacles. It looks like they rely on quite a bit of luck or they would crash into other obstacles while avoiding the obstacles they detect. Also, the world doesn't change much while you travel 10 meters, if you travel quickly, but your view of the world changes a lot when you turn. Each time they avoid an obstacle, the drone has not seen the nearest 10 meters in the new direction, and if an obstacle happened to be in the new path, it could very well already be too close to detect and avoid. The drone behaves very much like a panicky driver who veers into oncoming traffic to avoid hitting a squirrel.
Birds use a simpler approach: no 3d modeling; they just respond to relative speed of edges on their retina.