When autonomous drones are flying about without any knowledge of the environment, they have to fly slowly in order to avoid potential obstacles. That may not be a major problem in open fields, but it certainly is in crowded environments like urban areas or forests. Drones that are able to map their surroundings can better predict how to avoid obstructions, and therefore fly faster. However, the inherent fallibility of an IMU (inertial measurement unit) means those maps aren’t completely accurate and become less reliable over time.
Pete Florence at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) has come up with a better method, which abandons map-creation altogether. As a drone with Florence’s system flies, it collects snapshots of the readings gathered by its depth sensors. When it approaches an obstacle, it then compares its current readings with those it remembers from previous flights.
If it finds a similar recording, it then makes a decision about how certain it is that the new obstacle matches the old one. That judgement is based on how closely it matches the snapshot in memory, as well as how long ago it recorded it. If it’s uncertain about a match, it will fly slowly around the object to be safe. But, if it’s reasonably certain the obstruction has been encountered before, it can quickly avoid it by remembering how it did so in the past.