A team of MIT scientists has developed a low-cost sensor glove that allows artificial intelligence to identify objects by touch. If the research is successful, it will provide enhanced capabilities for next-generation robots and future biomechanical prostheses.
One of the areas of research is to collect as much information as possible about how people identify objects by touch. This will allow in the future to organize the process of machine learning to determine and analyze how the human hand works in the identification process, in particular, in determining the weight.
The laboratory prototype of the new device is an inexpensive knitted glove equipped with 550 pressure sensors. The glove is connected to a computer that collects and analyzes the data, which turns it into a video image of a “tactile map”. Further, the information is transmitted to the neural network, which classifies the images in order to extract pressure diagrams from them and compare them with specific objects.
During the research, 135, 000 video frames were filmed about 26 ordinary objects - cans of soda, scissors, tennis balls, spoons, pens and mugs. Then the neural network analyzed randomly selected frames until it learned to identify the objects depicted on them with 76% accuracy.
Another undeniable advantage of the MIT sensitive glove is that it costs only $ 10. For comparison, the cost of some of its counterparts with 50 sensors reaches several thousand dollars.