Force Perception for Rigid-Soft Finger Without Force Sensors: Theoretical Analysis, and Model Transfer

Abstract

Force perception is important for the manipulation of soft robotic hands. Multiple-direction interactions between fingers and objects occur predominantly at the fingertips during manipulation. Integrating physical multi-dimensional force sensors for soft fingertips poses stringent demands on the manufacturing process, material selection, and structural design. Therefore, we proposed a model-based and data-driven combined method to estimate the forces on the fingertip using the embedded liquid metal position and pressure sensors. This approach reduces the complexity of sensor system design and integration. In this work, we established the theoretical model of the rigid-soft finger, to generate a pre-training dataset after a pre-identification for theoretical model parameters. This dataset is used to pre-train a source network, and then transfer the source network with the real-world dataset to obtain the final force estimator with great generalization ability. The proposed method contributes to obtaining a better force estimator in fewer samples and biased datasets, effectively reducing the difficulty of data acquisition.