MagCilia Tracks 3D Forces and Distinguishes Six Surface Categories

A compact sensor built around flexible magnetic cilia reconstructed forces in three axes and distinguished six surface categories with 99.39% cross-validated accuracy. The same tactile unit also supplied tangential-force feedback for robotic gripper adjustment, according to an arXiv preprint describing the device, called MagCilia.

The sensor combines a flexible magnetic-cilia structure with a Hall sensor. Under multidirectional loading, the structure deforms and produces magnetic responses. Because the magnetic channels are coupled, the researchers used a regression method to convert the readings into estimates for Fx, Fy and Fz.

That method, causal history fusion regression, combines current magnetic-field measurements with their recent changes. The researchers also used quasi-static finite-element analysis to examine structural deformation and magnetic responses under loads from multiple directions.

In five-fold cross-validation grouped by calibration record, the system produced root-mean-square errors of 0.40 newtons for Fx, 0.57 newtons for Fy and 0.69 newtons for Fz. The corresponding coefficients of determination were 0.93, 0.90 and 0.92.

The robotic experiments included tangential-force-guided gripper adjustment. In another test, the tactile unit monitored loads along multiple axes under external perturbations. Both demonstrations used reconstructed force information from the compact tactile unit.

MagCilia also used frequency-domain features from its reconstructed force signals to distinguish six surface categories. It achieved 99.39% accuracy in three-fold cross-validation grouped by acquisition session. An online robotic demonstration additionally identified all six tested surfaces, although the supplied report does not state how many online trials were run or name the surface materials.

The reported approach combines force reconstruction and surface recognition in one compact tactile unit. The preprint presents the sensor as providing normal and tangential force information for grasping alongside surface information obtained through contact.

The supplied account does not state the sensor’s dimensions, load range, sampling rate, response latency, cilia materials or durability. The reported surface-recognition accuracy comes from grouped cross-validation across six categories, and the source does not describe broader testing on other surfaces or operating conditions. The work is presented as an arXiv preprint.