Before a robot can do anything useful, it has to make sense of the room it is standing in. Here is how AdamAI, our humanoid robot, looks around, learns new things, and recognizes them again later.

Seeing

AdamAI has a camera in his head. It gives him a live sense of what is in front of him, roughly how far away it is, and which side it is on. That live picture of the room is what the rest of him works from.

Learning by being shown

The part people find surprising is that you do not program AdamAI to know an object, and you do not sit him down for a long training session. You show him. He takes a look, remembers what the thing looks like, and from then on he can pick it out again. Teaching him something new takes seconds, not hours.

Scanning a room

To fill his memory, AdamAI does a room scan. He slowly turns his head across the space, notices the things around him, and then works through them one at a time: he centers on each object, takes a good look to be sure of what it is, and, if it is new to him, learns it and says its name out loud. Doing it one object at a time keeps him honest, so what he saves is really the thing he is naming. Every object he learns is filed with a little picture of exactly what he saw.

One memory for everything

The same idea runs through the rest of him. AdamAI keeps a single memory for what things look like, what has happened to him, and what he knows how to do, so his brain can look all of it up when it plans what to do next.

Teaching a robot to know the world usually sounds like a giant data and training problem. With AdamAI it feels more like taking notes: see it, save it, look it up later. He learns in seconds and it all runs on his own hardware.

More from the series at coburnicusmind.com.