A child exploring an unfamiliar apartment crawls from object to object and remembers what is visible from each corner. This is how the brain builds a mental map — an internal model of space. The gradCSCG algorithm works on the same principle: it receives a video stream from a camera and independently draws a map of the room, relying only on the changing images. Previously, machines needed hints — for example, pre-labeled objects in images. The new method does without them.

The key idea is a division of labor between two components. The first part of the program compresses visual scenes into a short code called a spectrum — a set of characteristic features of the scene, a kind of 'fingerprint' of what the camera sees. The second part, gradCSCG itself, links these spectral codes to the traveled path, recovering the topology: that is, the structure of space and the relationships between places. Both parts are trained simultaneously: recognition errors force topology corrections, and the refined topology in turn helps better recognition. This resembles two experts tuning each other, learning from one another.

To keep the system on track, mechanisms are built in that suppress entropy — here entropy means a measure of uncertainty or confusion in the algorithm's estimates. If entropy grows, the algorithm risks 'getting lost' in its own hypotheses. Additionally, there is protection against 'falling asleep' in monotony: when the image does not change for a long time, the system might stop refining the model. Special mechanisms maintain active learning, preventing the algorithm from settling on a wrong but stable solution.

An unexpected result: even if at the same point in space the camera showed at different times a handwritten digit '3' or a shapeless blotch, the algorithm confidently determined that it was the same point. It learned to ignore irrelevant changes in visual appearance, focusing on the topological identity of the place. This recovers the Standard Model — a typical plan of the location, a generalized representation of how the room is arranged.

The same principle works in the grid neurons in our brain — special cells in the entorhinal cortex that activate when an animal or human crosses the nodes of an imaginary hexagonal grid covering space. For their discovery, the Nobel Prize in Physiology or Medicine was awarded in 2014. This analogy is not accidental: gradCSCG reproduces a key mechanism of biological navigation, but in an artificial system.

The technology promises the emergence of assistant robots that can explore unfamiliar floors and rooms without GPS, like the droid BB-8 from the 'Star Wars' universe — a compact explorer that navigates unknown environments deftly, relying only on its sensors.