Member-only story
A Robot Pictured Its Body as a Cloud
Inside Columbia’s self-modeling robots: a 1.1-megabyte fog that caught a dead motor, and a one-camera successor that had to learn where it can’t see itself.
Ask a neural network what it thinks it looks like and don’t expect a portrait. When Columbia engineers finally got their robot arm’s self-image onto a screen, Hod Lipson described it as “a sort of gently flickering cloud that appeared to engulf the robot’s three-dimensional body.”
That cloud is the most honest picture of a machine self I know of. It isn’t a drawing of the robot. It’s what a question looks like when you ask it everywhere at once.
Questions like that have two ways to go wrong. You can believe you’re somewhere you aren’t, which is exactly how this robot caught its own broken motor. Or you can conclude you’re nowhere, and that’s where the lab’s next robot kept sliding during training: its cheapest early guess was that it wasn’t there at all.
A body made of answers
Most robots get their bodies from engineers. Someone draws the arm in CAD, writes the joint equations, and loads both into a simulator so the machine can rehearse a move before making it.
