Tactile Object Recognition Using Fluid-Type Sensor and Deep Learning

Tactile data is essential in object perception. Based on such data, various objects can be recognized and differentiated. The fluid-type tactile sensor employs a colored fluid captured between an elastic skin and a transparent plate where the color intensity of the fluid observed by a camera is directly related to the deformation of the skin. In the present work, the fluid-type sensor is restructured by employing an optimization method that is sensitive to skin deformation. Moreover, the implementation of non-opaque skin is proposed to reveal surface color features in addition to the shape of the object in tactile images. A CNN neural network based on transfer learning and dropout layer was implemented to classify objects based on their tactile images, which is a combination of shape of the objects and their surface color features. Although the number of training samples is small, an accuracy of 94.2% is achieved. The proposed sensor is 3D printable and can be fabricated at a low cost.