Last few years, I worked together with ESA and MIT on an experiment, in which a drone learns to see in zero gravity. Yesterday, ESA presented our joint work at the International Astronautical Congress in Mexico. In the experiment, the drone started to fly based on stereo vision with two cameras, but learned online to also see distances with a single camera. Already after a few minutes, the drone was able to predict distances with a single camera, which would allow it to keep navigating if one camera broke down. This type of “self-supervised learning” is very promising for future robot exploration in space.
We have also extensively tested the same learning algorithms on a drone on earth (see this article), but on earth it is easier to see distances, since gravity provides a nice reference frame and adds a lot of structure to the environment. In zero gravity, a drone can move in a full 6 degrees of freedom, so it can look at objects in many different ways.
See the news on the ESA page.