Source-linked AI summary
MagCilia: A Compact Magnetociliary Tactile Sensor with 3D Force Sensing for Robotic Contact Perception and Grasping Feedback
Yu Feng, Hao Wu, Haotian Guo, Haoming Liu, William Su, Jingxiang Guo, Jiankun Li, Masayoshi Tomizuka, Wen Jung Li, Jianshu Zhou
TL;DR
Magnetic-cilia deformation couples mechanical responses across magnetic channels, complicating quantitative 3D force reconstruction from contact loads. MagCilia uses a magnetic-cilia skin and Hall sensor, with CHFR combining current magnetic readouts and recent changes to reconstruct three force components. Calibration reconstruction yielded RMSEs of 0.40, 0.57, and 0.69 N for Fx, Fy, and Fz; session-grouped three-fold classification reached 99.39% accuracy, and robotic grasping feedback was demonstrated.
Problem
Magnetic-cilia deformation couples mechanical responses across magnetic channels, complicating quantitative 3D force reconstruction from contact loads.
Source: §I · p. 2Method
MagCilia uses a magnetic-cilia skin and Hall sensor, with CHFR combining current magnetic readouts and recent changes to reconstruct three force components.
Source: §I · p. 2, B. Causal History Features · p. 3Results
Calibration reconstruction yielded RMSEs of 0.40, 0.57, and 0.69 N for Fx, Fy, and Fz; session-grouped three-fold classification reached 99.39% accuracy, and robotic grasping feedback was demonstrated.
Source: §I · p. 2, A. 3D Force Reconstruction · p. 5, C. Classification Results · p. 7Takeaways & Limitations
The same tactile unit supports force reconstruction for grasping feedback and uses the temporal structure of reconstructed forces to distinguish surface categories.
Source: §I · p. 2
Abstract
from arXiv · showhide
Robotic grasping and surface exploration benefit from simultaneous measurement of normal and tangential forces and from surface information obtained through contact. Here, we present a compact magnetociliary tactile sensor (MagCilia) that combines a flexible magnetic-cilia structure with a Hall sensor for 3D force sensing. Quasi-static finite element analysis is used to investigate structural deformation and magnetic responses under multidirectional loading. To reconstruct forces from the coupled magnetic channels, we propose causal history fusion regression (CHFR), which combines current magnetic-field measurements with their recent changes. Five-fold cross-validation grouped by calibration record yields root-mean-square errors of 0.40, 0.57, and 0.69 N for Fx, Fy, and Fz, respectively, with corresponding coefficients of determination of 0.93, 0.90, and 0.92. Robotic experiments demonstrate tangential-force-guided gripper adjustment and multi-axis load monitoring under external perturbations. Frequency-domain features of the reconstructed forces distinguish six surface categories with 99.39% accuracy in three-fold cross-validation grouped by acquisition session. An online robotic demonstration additionally identifies all six tested surfaces. These results demonstrate 3D force reconstruction, grasping feedback, and surface recognition using a single compact tactile unit.
I. INTRODUCTION
MagCilia addresses the challenge of reconstructing 3D contact forces when magnetic-cilia deformation couples the magnetic channels. It combines a compact sensor and causal-history force reconstruction with robotic grasping feedback and surface recognition using the same tactile unit.
Motivation
- Simultaneous normal and tangential force measurement provides a fuller description of contact, where compression, object weight, perturbations, and relative motion can produce loads.Source: §I · p. 1
- Coupled cilia, membrane, and support deformation lets a force along one axis affect multiple magnetic channels, limiting direct channel-by-channel 3D force reconstruction.The introduction frames joint reconstruction as necessary because the magnetic responses reflect both mechanical deformation and cross-channel coupling.Source: §I · p. 2
Contribution
- MagCilia combines a magnetic-cilia skin, PDMS buffer, and Hall sensor; CHFR uses current magnetic readouts and causal history features to reconstruct three force components.The reported calibration RMSEs are 0.40, 0.57, and 0.69 N for Fx, Fy, and Fz, respectively.Source: §I · p. 2
Robotic applications
- Grasping experiments use reconstructed forces to adjust gripper opening under external loading, while surface scanning uses their temporal structure to distinguish categories.Both robotic applications use the same tactile unit.Source: §I · p. 2
II. MAGCILIA STRUCTURE AND SENSING MECHANISM
MagCilia stacks a magnetic-cilia skin over a deformable PDMS buffer and reads the resulting magnetic-field changes with a Hall sensor. Loading deforms the cilia and support in coupled ways, motivating joint reconstruction from the three magnetic channels.
A. Sensor Structure and Fabrication
- The magnetic-cilia skin contacts the object, the PDMS buffer supports and deforms with it, and the Hall sensor measures resulting magnetic-field changes.A housing supports and aligns the assembled layers.Source: A. Sensor Structure and Fabrication · p. 2
B. Magnetic Vector Response
- Normal loading primarily compresses cilia and support, while tangential loading bends cilia and displaces the base membrane with support-layer shear.Oblique loading combines compression, bending, and shear, jointly affecting magnetic-field output.Source: B. Magnetic Vector Response · p. 2, Figure 2 · p. 2
- The magnetic model accounts for changes in material position and orientation, while state-dependent cross-channel sensitivities motivate joint mapping from three magnetic channels to three force components.The model assumes relative permeability µr = 1, initial magnetization normal to the skin, and conservation of magnetic moment per reference volume.Source: B. Magnetic Vector Response · pp. 2–3
C. Finite Element Analysis
Quasi-static finite element analysis examines magnetic responses under normal and multidirectional tangential loading. The simulations show direction-dependent coupled responses, supporting use of the complete magnetic-field vector for calibration.
C. Finite Element Analysis
- The analysis uses a 5 µm normal pre-indentation and applies tangential displacements up to 2.5 µm along ±x, ±y, 45°, and 135° to compare directional shear responses.Source: C. Finite Element Analysis · p. 3
- The finite element model uses nearly incompressible neo-Hookean layers, a fixed bottom surface, and loading through a rigid top plate.Model dimensions are consistent with the magnetic skin and buffer layer.Source: C. Finite Element Analysis · p. 3
- Different loading directions produce distinct, coupled magnetic responses: normal indentation primarily changes Bz, orthogonal shear reverses the dominant transverse response with direction, and oblique shear changes both transverse components.The simulations motivate calibration using the complete magnetic-field vector but do not establish a unique inverse mapping from magnetic field to force.Source: C. Finite Element Analysis · p. 3
III. 3D FORCE CALIBRATION AND RECONSTRUCTION
The calibration pairs synchronized 3D reference forces with magnetic readouts across multidirectional loading. Causal history features augment the current readout with recent changes that can be computed from current and past samples.
A. Synchronized Data Acquisition and Organization
- Fig. 3 pairs normal indentation and x/y shear deformation with magnetic responses under orthogonal or oblique shear.The caption identifies the deformation and corresponding response panels for these loading modes.Source: Figure 3 · p. 3
- Five loading records provided 304,870 synchronized magnetic-field and reference-force sample pairs across varied force directions and magnitudes.The setup uses a reference force sensor and handheld indenter; Fig. 4 depicts the sensor installation, contact loading, and synchronized recording.Source: A. Synchronized Data Acquisition and… · p. 3, Figure 4 · p. 4
B. Causal History Features
- Causal history features augment the current magnetic readout with historical differences over 40 ms and 120 ms intervals.Historical samples use the latest timestamped data point no later than the target time, so features use only current and past information and can be computed point by point during acquisition.Source: B. Causal History Features · p. 3
- The two historical difference vectors preserve all three magnetic channels and describe finite changes, not force derivatives or additional mechanical measurements.Together with the current three-axis readout, these vectors form the causal history input.Source: B. Causal History Features · pp. 3–4
C. 3D Force Reconstruction
CHFR maps causal history features to three contact-force components by averaging predictions from two parallel regression branches. Its evaluation holds out complete calibration records and reports complementary force-reconstruction metrics.
C. 3D Force Reconstruction
- CHFR averages 3D force predictions from parallel Extra Trees and kernel-approximation ridge-regression branches using the same inputs.The model is illustrated in Fig. 5; the branches use local partitions and a smooth nonlinear approximation, respectively.Source: C. 3D Force Reconstruction · p. 4, Figure 5 · p. 4
- Five-fold cross-validation holds out one complete calibration record per fold and retains each test record in full.Training randomly samples up to 100,000 points for Extra Trees and up to 70,000 for kernel regression per fold; model selection uses the same record-level grouping.Source: C. 3D Force Reconstruction · p. 4
- Pooled out-of-fold predictions are evaluated with R2, RMSE, and MAE for each force component.Record-level grouping tests transfer between loading sequences, where adjacent samples can share contact states and overlapping histories.Source: C. 3D Force Reconstruction · p. 4
- The calibrated model provides three force components for contact monitoring and grasping feedback.RMSE and MAE quantify force errors in newtons, while R2 measures agreement relative to reference-component variation.Source: C. 3D Force Reconstruction · p. 4
A. 3D Force Reconstruction
Across five held-out calibration records, the reconstructed force components show strong agreement with reference forces. Dynamic-touch records also show tracking of normal and tangential force variations under repeated, lateral, and combined loading.
A. 3D Force Reconstruction
- R2 was 0.93 for Fx, 0.90 for Fy, and 0.92 for Fz; RMSE was 0.40, 0.57, and 0.69 N, respectively.Fig. 6 summarizes pooled out-of-fold predictions from five held-out calibration records; the reported MAE values were 0.25, 0.36, and 0.45 N for Fx, Fy, and Fz, respectively.Source: A. 3D Force Reconstruction · pp. 4–5, Figure 6 · p. 5
B. Dynamic Touch Response
- For one repeated-contact record, RMSE was 0.05 N for Fx, 0.07 N for Fy, and 0.12 N for Fz.The reconstructed normal force followed contact establishment, loading, and unloading in that record.Source: B. Dynamic Touch Response · p. 5
- During combined dynamic loading, predictions followed all three force components, with RMSE ranges of 0.05–0.10 N for Fx, 0.07–0.12 N for Fy, and 0.12–0.19 N for Fz.The tested interactions included repeated indentation and lateral loading, where tangential predictions tracked sign and magnitude changes.Source: B. Dynamic Touch Response · p. 5
V. DEMONSTRATION ON ROBOTIC GRASPING
MagCilia’s reconstructed forces guide gripper adjustment during loading and handover, while stepwise loads produce distinct tangential-force plateaus. The experiments also examine dynamic contacts with varying magnitudes, directions, and repetition rates.
A. CHFR-Based Force Feedback and Gripper Adjustment
- Sustained downward tangential loading triggers additional closure, while upward support or pulling triggers opening; return to baseline makes the gripper hold its opening.The controller uses a deadband and persistence criterion, with normal force monitoring contact and limiting further closure.Source: A. CHFR-Based Force Feedback and… · p. 5
- Fn rises from near 1.1 N during stable grasping to approximately 1.7 N under added loading and peaks at 3.91 N before returning near zero after release.The controller tightens the grasp during loading and opens it during upward support and object removal.Source: A. CHFR-Based Force Feedback and… · p. 5
B. Stepwise Loading During Grasping
- Mean Fx increases across three stable loading plateaus from 0.04 N before loading to 0.08, 0.11, and 0.17 N, while Fn rises from 0.32 to 0.48 N.Fy remains within a relatively small range during these incremental loads.Source: B. Stepwise Loading During Grasping · p. 5
- Dynamic-touch trials compare reference forces with predictions during contacts varying in magnitude, direction, and repetition rate.The figure includes fingertip loading and the reference-sensor setup.Source: Figure 7 · p. 6
C. External Perturbations in Different Directions
Grasp perturbations produce concurrent changes in normal and tangential forces. Active scanning then uses force signals from forward-and-return trials to capture surface-dependent variation across six categories.
C. External Perturbations in Different Directions
- During lateral perturbation, Fy spans −1.87–0.80 N; during vertical perturbation, Fx and Fy span −0.85–0.93 N and −0.73–0.77 N, while Fn peaks at 2.41 N.Perturbation direction changes the magnitudes and temporal responses across force components.Source: C. External Perturbations in Different… · p. 6
A. Active Scanning and Sample Composition
- The scanning dataset contains 328 forward-and-return trials across acrylic, fabric, paper, PLA, rubber, and sponge surfaces.MagCilia remains in contact as the robot scans each sample and reverses direction.Source: A. Active Scanning and Sample… · p. 6
- The median normal-force curve peaks at 2.50 N for paper and 0.73 N for sponge, while within-channel normalization emphasizes spectral power distribution rather than overall force level.The recognition pipeline converts reconstructed force traces into frequency-domain features for logistic-regression classification.Source: A. Active Scanning and Sample… · p. 6
B. Frequency-Domain Representation
The recognition pipeline represents reconstructed forces with normalized frequency-band features and evaluates classification across held-out acquisition sessions. Offline evaluation reports 99.39% accuracy, and the online demonstration identifies all six tested surfaces.
B. Frequency-Domain Representation
- Each trial is represented by 84 features formed from two scan directions, three force components, and 14 frequency bands, with per-channel band powers normalized by total power.Welch’s method estimates the spectra; normalization emphasizes relative spectral distributions and reduces dependence on overall signal amplitude.Source: B. Frequency-Domain Representation · p. 6
C. Classification Results
- The online robot scans six surface categories and correctly identifies all six tested surfaces after each forward-and-return scan.The workflow extracts frequency-band features and applies the trained classifier after each scan.Source: C. Classification Results · p. 6
- The predicted labels describe tested surface categories under acquisition conditions represented in the dataset; the online demonstration is not a separate accuracy estimate.The confusion matrix summarizes the session-grouped offline evaluation.Source: C. Classification Results · p. 7, Figure 11 · p. 8
- Across three-fold session-grouped evaluation, acquisition sessions were assigned so none appeared in both training and test sets within a fold; 326 trials were correctly classified, yielding 99.39% overall accuracy, 99.37% balanced accuracy, and 0.99 macro-averaged F1.Source: C. Classification Results · p. 7, Figure 11 · p. 8
VII. COMPARISON WITH RELATED WORK
Prior magnetic tactile work spans magnetized films and magnetic-cilia designs for multidirectional force or contact sensing. MagCilia combines one magnetociliary unit and one Hall sensor with robotic force feedback and surface recognition; the comparison is functional rather than a direct accuracy ranking.
VII. COMPARISON WITH RELATED WORK
- Magnetic tactile studies include self-decoupled normal and tangential sensing with sinusoidal films, three-component sensing with orthogonal films, and cilia-based sliding, weak tangential, flow-direction, or texture sensing.The cited examples include,, and –.Source: §VII · p. 8
- MagCilia uses one magnetociliary contact unit and one Hall sensor to reconstruct Fx, Fy, and Fz, then applies those signals to grasp feedback and frequency-feature surface recognition.The studies use different tasks and metrics, so the comparison is functional rather than a direct ranking of force accuracy.Source: §VII · p. 8
VIII. CONCLUSION
- Grouped five-fold cross-validation reports RMSEs of 0.40, 0.57, and 0.69 N for Fx, Fy, and Fz; future work targets composite and cilia geometry optimization and Hall-sensor arrays.The stated aims are improved force-reconstruction accuracy and spatial resolution.Source: §VIII · p. 8