Research on Data Construction and Classification of Deformable Objects Grasped by Soft Hand with Multi-source Information Fusion

The grasping of deformable objects poses a challenge to traditional rigid robots. Meanwhile, soft robots have advantages in robust grasping due to their inherent compliance. Data-driven methods have attracted the attention of many researchers, but the number of tactile time series datasets for soft robots is currently scarce. This paper utilizes a gripper containing three soft fingers controlled by 6 motors to grasp 8 deformable objects with different shapes, textures and softness. The tactile sensor array embedded in each finger can change the output voltage according to the deformation of the object and the finger which provide 8-channel tactile information, and finally obtain 24-channel tactile data of three fingers. At the same time, the position and torque data of the motors are collected. Finally, a soft-hand grasping deformable object dataset TAM (Tactile and Motor) is constituted. TAM has 37 channels in total, including time stamp, 24 haptic data channels and 12 motor data channels. The classification results of deep learning algorithms and other classic machine learning algorithms are compared on this dataset. Different sources of data are segmented and discussed to study the impact of haptic data and motor data on classification accuracy. Finally, we also put forward a vision for future work. The dataset will be open on a public platform for researchers to facilitate subsequent tasks such as object recognition of soft robots.