iRobot built over 100 models of dog poop to train Roomba's visual AI

Apparently Roomba owners who also own dogs have long encountered a problem: If the dog poops on the floor, their helpful robot vacuum rides right over it and smears it across the living room.

So iRobot decided to build a visual-detection model to help Roombas recognize, and avoid, dog poop.

According to iRobot's promo video, and this story in Techcrunch, this involved building over 100 physical models of dog poop and sending employees home to photograph them on their houses' floors:

"The glorious career of roboticists may not have been fully realized when we were sending people home and creating hundreds of models of poo," CEO Colin Angle recently told me. "Sending people around to photograph and create synthetic models of poo. I don't know how many tens of thousands of images of all different shapes and sizes of synthetic images were required, but this is not demo code, clearly." [snip]

"You imagine it, we probably attempted it to grow a large enough database with both real images, images of fake poo and synthetic images that were manufactured of poop to serve as a training model for our robot," Angle adds.

All of this leads us to Pet Owner Official Promise (P.O.O.P.), which guarantees a free return for the new j7+ if the Roomba runs into (and over) a poop problem. For the time being, however, iRobot is strictly adhering to the old adage about letting yellow mellow. "We can't do pee," says Angle. "It has to have some 3D aspects to it."