Source-linked AI summary
VecTile: A Low-Cost Flexible Tactile Array for 3-DoF Force Sensing
Jian Hou, Xin Zhou, Adam J. Spiers
TL;DR
Flexible, low-cost tactile arrays often measure normal force but not shear, despite shear’s relevance to contact. VecTile combines piezoresistive sensing with silicone nubs to recover three-axis force at each taxel. Hardware tests report 0.39 N normal and 0.16 N shear RMSE over the stated force ranges.
Problem
Flexible capacitive and piezoresistive arrays conform to curved surfaces but do not provide shear information, leaving accessible flexible arrays with local 3-DoF force measurements underexplored.
Source: §I · p. 1Method
VecTile combines a piezoresistive array with passive silicone nubs that convert normal and shear forces into distinguishable quadrant pressure readings.
Source: §I · p. 1Results
0.39 N normal RMSE and 0.16 N shear RMSE were achieved over 1–5 N normal force and up to 1.5 N shear.
Source: §VI · p. 8Takeaways & Limitations
The sensor and readout board cost less than 35 USD and can be produced in-house without expensive equipment or tools.
Source: §I · p. 1Sensitivity loss above roughly 5 N per taxel bounds the force range, and hysteresis, drift, and long-term durability remain uncharacterized.
Source: §VI · p. 8
Abstract
from arXiv · showhide
Tactile sensing provides robots with contact information difficult to obtain from vision alone. Shear force is particularly important for understanding tangential contact, guiding corrective motions, and maintaining grasp stability, yet flexible and low-cost tactile arrays often measure only normal force. We introduce VecTile, a low-cost, 6 mm thin, flexible tactile array capable of three-axis force sensing. VecTile combines a piezoresistive layer with passive silicone nubs that convert normal and shear forces into distinguishable pressure patterns. The sensor can be fabricated in-house without specialized equipment. The complete system costs less than 35 USD for a 4x4 array. We validate the sensing mechanism through finite element analysis and hardware experiments. In controlled single-taxel tests over 1-5 N normal force and up to 1.5 N shear force, VecTile achieves root-mean-square errors of 0.39 N and 0.16 N for normal and shear forces. A VecTile calibrated while flat retains accuracy when curved, while a five-minute manual calibration is sufficient for practical use. Two applications demonstrate its utility. First, shear readings are sensitive enough to stop a sharp scalpel from puncturing a balloon after contact. Second, the sensor substantially improves Diffusion Policy-guided power-plug insertion. All designs and code will be released upon paper acceptance. Project website: https://jiannnh.github.io/VecTile/.
I. INTRODUCTION
Flexible tactile arrays often lack local shear sensing, although shear helps characterize tangential contact and grasp stability. VecTile combines a flexible piezoresistive array with silicone nubs to provide local three-axis force sensing at low cost.
I. INTRODUCTION
- Shear complements normal-force readings by revealing tangential loading relevant to grasp stability, slip reasoning, and contact regulation.Flexible skins also need to conform to non-planar surfaces, while low cost and repairability matter because direct contact exposes sensors to wear and damage.Source: §I · p. 1
- Flexible capacitive and piezoresistive arrays conform to curved surfaces but lack shear sensing, leaving accessible local 3-DoF flexible arrays underexplored.Vision-based and magnetic sensors can recover richer contact information but are rigid.Source: §I · p. 1
- VecTile is a 6 mm-thin, flexible 4×4 array with 16 local 3-DoF taxels at 10 mm pitch, combining a piezoresistive array with passive silicone nubs.The nubs enable each taxel to read a three-dimensional force vector.Source: §I · p. 1
- The sensor and readout board cost less than 35 USD, and the sensor can be produced in-house without expensive equipment or tools.The stated cost includes a 25 USD Arduino Nano, which could be replaced with a lower-cost microcontroller and ADC circuit.Source: §I · p. 1
- 0.39 N normal RMSE and 0.16 N shear RMSE were achieved over 1–5 N normal force and up to 1.5 N shear per taxel.The paper also assesses flat, convex, and concave configurations and a short manual calibration process.Source: §I · p. 1
II. RELATED WORK
Prior multi-axis tactile sensors use vision or magnetic transduction, while conformable capacitive and piezoresistive arrays commonly measure normal pressure alone. VecTile uses raised nubs to convert shear into differential pressure, building on a mechanism explored in earlier work with fabrication trade-offs.
II. RELATED WORK
- Vision-based sensors estimate force from elastomer deformation but require rigid housings for cameras, illumination, and optical paths; magnetic skins complicate large-area integration and bending.eFlesh [10] and ReSkin [15] place magnetic skins over magnetometer boards.Source: §II · p. 2
- Thin capacitive and piezoresistive matrices form conformable tactile skins but primarily measure normal pressure, [13]; CoinFT recovers a six-axis wrench from capacitive layer displacement yet remains a rigid, single sensing unit rather than a distributed skin [19].Source: §II · p. 2
- VecTile uses raised nubs to convert shear into differential normal pressure, a mechanism also explored by Huh et al. and Yu et al..The cited prior designs involve spin-coating silicone on PCB or material coating, etching, and bonding, which the authors contrast with practical, cost-effective in-house fabrication.Source: §II · p. 2
III. SENSOR DESIGN
VecTile uses a scanned flexible-PCB array of four piezoresistive quadrants beneath each silicone nub. Balanced quadrant readings encode normal load, while across-quadrant imbalance encodes shear; finite-element modeling tests this mechanical transduction design.
A. Sensing Element and Readout Electronics
- Each 4×4 taxel contains four piezoresistive quadrants, forming an addressable 8×8 row–column matrix scanned by driving one column and sensing one row.Compression lowers the silver-fiber layer’s resistance across electrode pairs; the design places both electrodes of each quadrant on one FPC, avoiding precision alignment between electrode layers.Source: A. Sensing Element and Readout… · pp. 2–3
- An added LM358 transimpedance amplifier with non-activated columns held at virtual ground addresses early saturation seen with a simple voltage-divider readout under simultaneous nub loading.The selected quadrant’s ADC output is approximately VdriveRref/Rij; compression lowers Rij and raises output, while a potentiometer adjusts readout gain.Source: A. Sensing Element and Readout… · p. 3
1) Working Principle:
- Each silicone nub covers four quadrants: normal force raises all four readings together, while shear tilts the nub and shifts load between opposing quadrants.The design encodes normal force in the balanced response and shear in the imbalance.Source: Figure 2 · p. 2, 1) Working Principle · p. 3
2) Finite Element Model:
- The finite-element model tests four nub heights from 0.6 to 3.0 mm under 1–5 N normal preload followed by 1 N positive-x shear, assuming in-plane symmetry.The model uses Dragon Skin 30 with a near-incompressible Mooney–Rivlin material model and fixed corner patches on the nub cap.Source: 2) Finite Element Model · p. 3
- The model defines Σ as the sum of contact forces over the nub’s two x-axis halves and D as their difference, measuring contact imbalance.These readings are the model’s basis for assessing how normal and shear loading are encoded.Source: 2) Finite Element Model · p. 4
3) Mechanical Separability of Forces:
Simulations show that normal and shear loads primarily affect different readings, supporting force inversion over the sampled positive-x range. Increasing nub height raises simulated shear gain while preserving most transmitted normal force.
3) Mechanical Separability of Forces:
- 86–90° separation between normal- and shear-response directions across tested states supports inversion from (Σ, D) to (Fz, Fx) within the sampled positive-x range.Normal-force increments act mainly through Σ, while shear increments act mainly through D.Source: 3) Mechanical Separability of Forces · p. 4
4) Simulated Effect of Nub Height:
- At Fz = 5 N, Σ falls from 5.00 N for a 0.6 mm nub to 4.44 N for a 3.0 mm nub, while simulated differential response increases with nub height.Taller nubs require more indentation to support the same normal force, but most applied normal force reaches the nub footprint for every tested height.Source: 4) Simulated Effect of Nub Height · p. 4
C. Sensor Fabrication
VecTile is assembled by molding a silicone layer, inserting the FPC and resistive layer, and bonding the stack while preserving local compliance. The sensing layer plus readout costs $34.67 per unit at the stated order quantities, and sensor characterization follows.
C. Sensor Fabrication
- The fabrication sequence molds and cures a silicone layer, inserts the FPC and resistive layer, then injects silicone through openings to bond the stack.The molded layer cures for 16 hours; the final injected silicone bond cures for approximately 8 hours under a 200 g load.Source: Figure 4 · p. 4
- Cyanoacrylate bonds the resistive layer to the FPC between sensing regions, while injected silicone bonds it to the molded layer without constraining local deformation.Bonding materials stay outside electrode areas, and the assembly has no rigid backing, preserving flexibility.Source: C. Sensor Fabrication · p. 4
- The sensing layer plus readout costs $34.67 per unit and could cost under $15 with a cheaper microcontroller.The projected lower cost is $14.67 with an approximately $5 ATmega328P development board.Source: C. Sensor Fabrication · pp. 4–5
IV. SENSOR CHARACTERIZATION
- The physical sensor evaluation tests how nub height and curved mounting affect raw response and decoded three-axis accuracy.The characterization section follows the fabrication description.Source: §IV · p. 5
A. Transduction of the Resistive Layer
The resistive layer was characterized separately to establish its response to normal loading and select a working range for whole-sensor tests.
A. Transduction of the Resistive Layer
- The FPC and resistive fabric were mounted without silicone, and a guide plate aligned a loading pin to each taxel for 20 loading cycles per taxel.The pin covered four quadrants, so force and summed counts were divided by four for per-quadrant reporting.Source: A. Transduction of the Resistive Layer · p. 5
- ADC response becomes less sensitive near 1.25 N per quadrant, corresponding to approximately 5 N per taxel, so whole-sensor characterization used 1–5 N per nub.This is not a hard force limit; additional force is simply harder to resolve accurately.Source: A. Transduction of the Resistive Layer · p. 5
- Fig. 5 pairs the force-response measurement with a per-quadrant ADC plot whose shaded envelope spans 64 quadrants.The quadratic fit covers Fz ≤1.25 N.Source: Figure 5 · p. 5
B. Data Collection Setup
Controlled robot-arm tests compared four nub heights under normal and tangential loading, including curved mounting of the tallest design.
B. Data Collection Setup
- Four flat builds with nub heights of 0.6, 1.0, 2.0, and 3.0 mm were tested; the 3.0 mm build was also tested on convex and concave mounts with R = 85 mm.An ATI Mini40 between the rigid probe and UR5e wrist recorded ground-truth forces.Source: B. Data Collection Setup · p. 5
- Each controlled interaction pressed perpendicularly to a target preload, held for 2 seconds, moved laterally by 2 mm, held again, then released.The procedure began with a non-contact baseline to zero VecTile and the force/torque sensor; no slipping was observed.Source: B. Data Collection Setup · p. 5
- Across five preload levels, eight shear directions, and 16 nubs, each sensor yielded 640 interaction sequences and 348,800 data points were collected.Data were saved at 86.5 Hz.Source: B. Data Collection Setup · p. 5
C. Force Reconstruction Across Nub Heights
Taller nubs improve shear-force reconstruction while retaining similar normal-force accuracy; curvature modestly increases error when using a decoder fitted on the flat sensor.
C. Force Reconstruction Across Nub Heights
- At 1 N transmitted shear, median projected differential rose from 34.2 counts for the 0.6 mm nub to 101.6 counts for the 3.0 mm nub, while normal ADC response changed little with height.Greater nub height lowers stiffness but significantly improves shear sensitivity.Source: C. Force Reconstruction Across Nub… · pp. 5–6
- Each taxel's quadratic decoder maps its four baseline-subtracted quadrant readings to force components using ordinary least-squares-fitted coefficients.The evaluation withheld opposing shear-direction pairs across four folds and normal-force levels.Source: C. Force Reconstruction Across Nub… · p. 6
- The 3.0 mm nub achieved 0.16 N shear RMSE versus 0.29 N for the 0.6 mm nub, while the four designs had similar normal RMSE.The wider shear-count range improved reconstruction without sacrificing normal-force accuracy.Source: Table II · p. 6, C. Force Reconstruction Across Nub… · p. 6
- Using the decoder fitted on the flat 3.0 mm sensor, curved mounts increased average normal and shear RMSE by 0.05 N and 0.06 N, respectively.The curved tests used mounts with R = 85 mm.Source: C. Force Reconstruction Across Nub… · p. 6
D. Practical Calibration & Evaluation
VecTile’s brief manual calibration supports practical force estimation, and its shear channel enables a scalpel reflex that retracted before puncture in repeated trials. The section also evaluates multi-nub prediction and closed-loop manipulation demonstrations.
D. Practical Calibration & Evaluation
- A manually calibrated model achieved per-taxel RMSEs of 0.43 N for normal force and 0.22 N for shear force, with a median 25 ms lag to the F/T reference.Calibration used about 1,400 data points from 10 fast manual interactions per nub; testing used 240 data points.Source: D. Practical Calibration & Evaluation · p. 6
- Multi-nub predictions achieved RMSEs of 0.71 N for normal force and 0.43 N for shear force after manual calibration.Predictions summed per-taxel forces; the 624 recorded frames included loads up to 15.54 N and 5.83 N on the F/T sensor.Source: D. Practical Calibration & Evaluation · p. 6
V. SENSOR DEMONSTRATION
- Two closed-loop tasks used the manual fast-calibration method: a shear-sensitive scalpel reflex and visuotactile Diffusion Policy plug insertion.The scalpel task demonstrated retraction after contact detection without bursting the balloon.Source: §V · p. 6, A. Shear-Aware Scalpel–Balloon Reflex · p. 7
A. Shear-Aware Scalpel–Balloon Reflex
- In 20 of 20 trials, the shear-triggered reflex retracted the scalpel before puncture, while the matched normal-only threshold did not trigger before the balloon burst.The robot retracted when any taxel’s shear-force change exceeded 0.1 N; the initial burst trial measured approximately 0.5 N summed shear change along the loading direction.Source: A. Shear-Aware Scalpel–Balloon Reflex · p. 7
B. Visuotactile Plug Insertion
The plug-insertion evaluation compares vision-only, normal-tactile, and full 3-DoF tactile policies at unseen socket locations. Success rose across those conditions, while shear readings supported corrective adjustments and contact-state interpretation.
B. Visuotactile Plug Insertion
- The task required aligning and inserting a UK 3-pin plug held by the gripper, with wrist and fixed scene cameras providing two views.The setup is shown in Fig. 8(B.i).Source: B. Visuotactile Plug Insertion · p. 7, Figure 8 · p. 7
- The evaluation used 84 teleoperated demonstrations across six socket locations and tested 25 rollouts per policy at locations unseen during training.Each location contributed seven direct and seven corrective trajectories; test locations remained within the training region’s spatial range.Source: B. Visuotactile Plug Insertion · p. 7
- Each policy combined visual features with tactile input, end-effector pose, and gripper width, and predicted action chunks executed at 10 Hz.The three conditions were vision only, vision with normal-only tactile input, and vision with full 3-DoF tactile input.Source: B. Visuotactile Plug Insertion · p. 7
- Plug-insertion success improved monotonically across vision-only, normal-only tactile, and full 3-DoF tactile conditions.The table reports performance at unseen locations, but the supplied passages do not give its cell values.Source: B. Visuotactile Plug Insertion · p. 7, Table III · p. 7
- During a typical successful rollout, tactile patterns marked contact transitions, including asymmetric alignment forces and opposing shear after seating, which cameras occluded and normal-only input lacked.The sequence included an initial failed direct insertion, a finding-hole step, alignment, final descent, seating, and release.Source: B. Visuotactile Plug Insertion · p. 7
VI. CONCLUSION
VecTile uses passive silicone nubs to recover three-axis force at each taxel in a flexible, low-cost array. The conclusion reports force-estimation results and identifies force-range and durability-related limits.
VI. CONCLUSION
- The 3.0 mm design achieved RMSEs of 0.39 N for normal force and 0.16 N for shear force over 1–5 N normal force and up to 1.5 N shear force.A flat-calibrated decoder transferred to convex and concave mounts with an average RMSE increase of 0.06 N, and a few minutes of manual calibration proved sufficient for practical use.Source: §VI · p. 8
- VecTile is a flexible 4 × 4 array that uses passive silicone nubs to convert shear into pressure differences across four piezoresistive quadrants.The complete unit costs under 35 USD, or under 15 USD with a cheaper microcontroller.Source: §VI · p. 8
- Sensitivity loss above roughly 5 N per taxel bounds the force range, and hysteresis, drift, and long-term durability remain uncharacterized.The authors propose exploring other elastomers or nub geometries and systematically quantifying time-dependent effects.Source: §VI · p. 8