A Learning-Based Multisensor Fingertip Force Measurement Method With Fast Data Generation

Force measurements become important once contact has occurred and the object is held or explored by the hand. The uncertainty associated with the fingertip contact position and the soft finger contact torque can lead to inaccurate force measurement when using a model-based method. To address this issue, we utilized a learning-based method that leverages neural networks to concatenate data from tactile, position, and torque sensors to improve force measurement accuracy in soft finger contacts. Due to the requirement for a large volume of data to train the network, we propose a data collection method that is independent of the fingers. This method utilizes the physical model of the finger to rapidly generate a large dataset. The effectiveness of this data collection method is validated on real-world datasets. Comparing the model-based method and the data-driven method using solely tactile input, our proposed force measurement method achieves the lowest force errors in all three axes and magnitude. Particularly, in the
${x}$
-axis and force magnitude, the proposed method reduces the error by more than 50% compared to the model-based method. Furthermore, our method demonstrates effectiveness and robustness on unseen objects. This method is not only suitable for dexterous hands but also applicable to underactuated hands equipped with torque sensors and high-dimensional tactile.