What you see and where you are
How flies, humans, vacuum cleaners and self-driving cars explore and map the environment — and how this led to a Nobel Prize
Before GPS-based navigation systems became ubiquitous, if you had to drive to someone’s house for the first time you would call them up and get directions to their place. Getting to a destination this way often included some painful backtracking and sometimes an even more painful arguments with the others in the car. But, behind the scene (pun intended), a set of very high-level cognitive processes were hard at work to create a mental map of the environment and, at the same time, manage to determine the current position within that map.
But not all organisms navigate their world via internal representations. Try to wave a fly out of a window and, most likely, you will end up looking at the critter’s inability to realize that freedom is just a few centimeters away. Of course, insects are fully capable of spatial orientations, too, but they do so by using pre-defined exploration patterns such as spiraling around or random zig-zagging (these strategies are actually very successful, too, and are behind the first successful domestic robot ever sold, the Roomba vacuum cleaner).
Humans, however, need to be able to efficiently navigate a complex and sometimes unknown terrain full of constantly changing features. For this, a mental map of the environment has to be created in our heads, and constantly updated, while at the same time keeping track of the location within that map. This circular refinement between map and position also lies at the heart of many successful computational solutions called SLAM (for Simultaneous Localization and Mapping), and they are what make modern autonomous self-driving cars possible.
Here’s how SLAM algorithms work. Let’s pretend you are driving a car without GPS, as in the old days, and that your instructions read: “after getting on the main road, drive for about two kilometers and turn left after the red building”. You now have two pieces of information that create your first internal map, namely the distance (two kilometres) and the landmark (red building). You will now try to minimize the errors with respect to that internal map and your location in it, while at the same time taking new visual inputs into account. If for example, right after joining the main road you see a red building, you will probably not turn left, since that would imply that the distance information is completely wrong. On the other hand, as you get closer to the two-kilometer mark, any red building would be interpreted as the moment to turn and would thus reset your position. A self-driving car works the same way: it integrates imprecise positional information provided by the GPS, while using local visual landmarks detected by cameras and laser-based detection devices.
By the way, the medicine Nobel Prize in 2014 went to a trio of researchers who found the existence of a GPS-like system inside our human brains, responsible for creating cognitive maps that allow us to orient ourselves in space, while using feedbacks from our visual system — just like self-driving cars do.
