MIT researchers are working on a new drone piloting system that uses uncertainty to make sure they do not hit obstacles as they fly autonomously. The system is a bit complex but it is called NanoMap and it simply finds ways to move from point A to point B without crashing and manipulating random objects in its path.
Spectrum describes the system in detail but, basically, the drone takes a measure of depth by moving along a path. Whenever he takes a step and that he is about to move forward, he goes over the previous steps that could include information relevant to the current move. If he finds anything useful, he slows and evaluates the area and, if he finds any previous information, he continues to fly, avoiding obstacles.
This is important because current models require a drone to map its environment before becoming "confident" that it can handle the flights faster. This technique creates a map on the fly that allows the drone to handle uncertainty rather than being ready in all situations. In addition, this allows drones to sneak between pillars or trees and base their next movement on information collected on the fly and not on time. From the study:
In the tests, the researchers found that their modeling of uncertainty really began to pay off when the drift was much worse than 20 cm / s. Up to about 75 cm / s of drift, planning with NanoMap and the integration of uncertainty prevented the drone from crashing in 97 to 98% of cases. With a drift of more than 1 m / s, the drone was only 10% of the time, but it was three times more robust than the test without modeling uncertainty. The press release summarizes broadly:
If NanoMap did not model uncertainty and the drone drifted only to within 5% of the intended location, the drone would crash more than one. times every four flights. Meanwhile, when he took into account the uncertainty, the accident rate reduced to 2 percent.
The team is at MIT's CSAIL lab and is headed by Pete Florence in Russ Tedrake's lab. You can read the document here and read more here.