"iRobot hand-sculpted hundreds of fake dog turds out of Play-Doh and paint to provide training examples for its Roomba object-recognition system."
Colin Angle told IEEE Spectrum that iRobot created hundreds of artificial poop models using Play-Doh and paint, alongside real photographs and synthetic training images. The story illustrates how even a narrowly defined machine-learning task can require substantial effort to represent the relevant physical world in training data.
View Source