A deep neural network for tactile perception in open scenes

Tactile perception is important for the robots to understand their working environment. While in real-world applications, robots usually must face unexpected changes in external conditions, such as the re-installation of the robot end effector or the change of the installation location. Consequently, the collected tactile material data tend to vary to a certain extent, which brings great difficulties to the tactile perception. To handle this problem, different from the former studies of tactile perception in enclosed environments, this study focuses on the tactile material recognition task using robot electronic skin in open scenes. We construct a cross-batch tactile dataset to simulate open scenes and propose the multi-receptive field attention enhancement network (MRFE) to handle tactile material recognition. Compared with other machine learning algorithms, experiments show that the proposed method overcomes the problem of data drift caused by changes in posture, contact force, sliding velocities, exploratory motions, and assembly conditions.