The research team of Columbia University School of Engineering announced on July 13 that a robot they created can learn its whole body model from scratch for the first time without any human help. In the study published in science robotics, researchers described how their robots create their own kinematic model, and then use their own model to plan motion, achieve goals, and avoid obstacles in all cases. It will even automatically identify and compensate for damage to the body. The researchers placed the robot arm in a circle composed of five streaming media cameras, and the robot observed itself through the camera and swung freely. Just like a baby exploring itself in a hall mirror for the first time, the robot knows how its body moves in response to various motion commands by swinging and twisting. About three hours later, the robot stopped. Its internal deep neural network has completed the learning of the relationship between the robot motion and the volume it occupies in the environment. Howard Lipson, a professor of mechanical engineering and director of the creative machine laboratory at Columbia University, said that people really want to know how robots "Imagine" themselves, but people can't just spy on neural networks, because this is actually a "black box". The ability of robots to model themselves without the assistance of engineers is very important. It can not only save labor, but also let the robot follow up its own wear, and even detect and compensate for damage. The research team believes that this ability is critical because humans need autonomous systems to be more self reliant. For example, a factory robot can detect that something is not moving correctly and make compensation or seek help. Lipson explained, "self shaping is a primitive form of self-consciousness. If robots, animals or humans have accurate self models, they can play a better role in the world, make better decisions, and have evolutionary advantages." (Xinhua News Agency)