Researchers at the Artificial Intelligence Laboratory have developed a "tactile glove" with 548 sensors that can be used to identify objects, the weight of objects, and more.
The glove consists of a hand-shaped sensor cover and an ordinary knitted glove. It consists of a plain knit glove (yellow) and a hand-shaped sensor cover (black).
The sensor sleeve is arranged in two layers of 64 conductive lines, 32 in the horizontal and vertical directions, and a tension sensitive film (a film sensitive to vertical force) between the two conductive lines.
These lines intersect 548 points, each of which is a pressure sensor. When these points are pressed, the resistance of the film at the intersection becomes smaller and the electrode array can be perceived.
The output of the glove can be processed into a 32 x 32 array of grayscale pixels, where the color of each pixel represents pressure change, black represents low pressure, and white represents high pressure.
The researchers recorded the pressure map at seven frames per second. The pressure map collected by the sensor while using the glove allows the machine learning model to learn to identify the object, estimate the weight of the object, and distinguish between different hand postures.
To prove that the glove captures the different interactions between the hand and each object, the researchers used the recorded data for automatic object recognition. They show how a state-of-the-art deep learning model learns to re-identify 26 types of objects from collected pressure map data, which was originally designed for large-scale image classification.
Using only haptic data, the AI system recognizes objects with an accuracy of up to 76%. Experiments have also shown that a large number of pressure maps and their spatial resolution are key to successful identification of targets.
Next, the author uses gloves to pick up objects and proves that a similar deep learning model can estimate the weight of an unknown object. The results show that most of the objects weighing less than 60 grams can be accurately estimated.