A robot powered by an artificial intelligence that has been trained to tie knots could enable simple machines to perform intricate tasks with electrical wires or cables.
Robot arms can solder microscopic components or construct an entire car from thousands of parts, but most struggle to deal with flexible items like a piece of string because they move in unpredictable ways.
Tetsuya Ogata at Waseda University in Japan and his colleagues have now taught a robot with a pair of general-purpose arms, each equipped with two fingers, to tie a knot around a box as if it were being gift-wrapped. The robot is able to tie two kinds of knots: a bowknot, which is commonly used to tie shoelaces, and an overhand knot, a simple knot sometimes tied at the end of a rope to prevent fraying.
Advertisement
The team began by manually knotting a piece of rope dozens of times using the arms by remote control. The information recorded by the arms during those tests was then combined with data from an overhead camera and proximity sensors on the fingers and used to train a neural network to replicate the procedure.
During the training, the team used a rope with one half coloured red and the other half coloured blue to aid identification. When tasked with tying a bowknot using rope with coloured sections, the robot succeeded 95 per cent of the time. It also succeeded 90 per cent of time using a plain white rope, despite not having been trained with it. With overhand knots, the robot got it right 95 per cent of the time with a coloured rope and on 85 per cent of attempts with white rope.
Ogata says the team intends to improve the robot by adding a torque sensor to the arms, giving it information on how hard it is pulling the string and allowing it to tie tighter knots.
Reference:
