Researchers at Columbia University have created a robot that can learn and adapt to damage by observing itself—much like children, professional dancers and athletes do when they watch themselves in a mirror. This was made possible by a “kinematic self-awareness” system that lets it analyze videos of its movements and use them to adjust its subsequent actions.
To achieve this, the scientists used three neural networks: the first interprets video, the second converts the data into a model of moving parts, and the third trains the robot to plan movements in real-world conditions. The robot successfully completed tests, avoiding obstacles and compensating for damage by watching recordings of its actions.
The development opens up prospects for robots that can operate without constant human intervention, which is especially important for machines working in manufacturing, hazardous conditions or care settings. Perhaps soon a robot vacuum cleaner will stop bumping into furniture and start correcting its own path.
