Robot Masters Nunchucks After Learning like a Human Student
Advanced motor skills are essential for robots to physically coexist with humans. Cornell University researchers have shown a robot how to flip nunchucks, the martial arts weapon, to demonstrate an intuitive approach for teaching complex manual tasks. The group built a bionic hand and a motion-capture glove that can be used to teach the robot by demonstration, a popular method for skills requiring dexterity. Robot learning from human demonstration (LfD) has limitations in handling dynamic skills and compound actions. The Cornell researchers introduce a composite learning scheme that goes beyond LfD, and integrates robot learning from human definition, demonstration, and evaluation. First, the teacher explains each step of the trick using an intuitive symbolic flow chart called a Petri net. Then, they demonstrate the trick and evaluate their own performance after every attempt, creating data the robot uses to learn the movements required at each step and develop criteria for self-evaluation as it practices. It took the robot a matter of hours to learn how to spin the nunchuck around the back of its hand and catch it again. Since the approach is not task-specific, the researchers say it could help teach robots various complex, dynamic motor skills that could include assembling car interiors or fruit picking.