Scientists from the Laboratory of Robotics and Embedded Artificial Intelligence (AI) at Stanford University (USA) used sound to train robots, arXiv writes.
According to researcher Zei Liu, audio materials can contain a lot of useful data. Through an experiment, experts have proven that learning through sound can be more effective than without it.
During the study, scientists took a test robot and taught it various operations using sound. The machine was given four audio tracks to listen to - the sound of a bagel flipping in a pan, chalk being wiped off a blackboard, two sticky surfaces joining together, and dice falling from a glass.
The scientists demonstrated each task several hundred times and found that after sound training, the robot performed its tasks with 94 percent efficiency. With no sound - just a picture - the efficiency of the machine dropped to 27 percent.
„We can say with confidence that sound is the least studied mode of perception in robots,” said Dmitry Berenson, an associate professor of robotics at the University of Michigan, who reviewed the scientific work of his colleagues. The material states that to increase efficiency, robots must learn to distinguish external sound from the noise that their parts and components make during operation.