Multipoint Contact Detection for Rigid-Soft Finger Without Tactile Sensor, Using Domain Adaption Combined Network

Proprioception including position and force perception is particularly important in the grasp and manipulation of soft robots. However, limited by their size and fabrication, it is difficult to integrate a sufficient number of high-performance soft force/tactile sensors for the soft structure. The lack of sufficient sensors makes it difficult to recognize their own state and the external environment. To address this problem, a method of contact-position prediction by the actuator-pressure sensor and joint-angle sensor without physical force/tactile sensors is proposed in this work. In order to deal with the common problems of manufacturing differences between individuals, and characteristic changes caused by stress relaxation, a domain adaption combined network based on log correlation alignment is proposed. In our method, we use a complete contact-labeled dataset of one finger to predict other fingers with missing labels and use the latest unlabeled data to update the model after finger characteristic changes. Our method is verified by experiment, which can utilize small sample datasets with dataset bias, effectively improve the prediction accuracy in cross-individual and cross-time prediction, and greatly reduce the collection amount of labeled data.