VecTile: A Low-Cost Flexible Tactile Array for 3-DoF Force Sensing

VecTile on Fin Ray-type gripper fingers inserting a plug into a socket. Insets show the 4x4 taxel readings of each finger, colour for normal force and arrows for shear. Bottom left: a VecTile bent between two fingers.

Flexible tactile skins usually feel only how hard they are pressed. VecTile adds a passive silicone nub over every four piezoresistive cells, so one cheap, bendable array reports normal and shear force at all 16 taxels. We characterize the full sensor against a force/torque reference, taxel by taxel.

16

3-DoF taxels at 10 mm pitch

<6mm

thin and flexible

<$35

sensor + readout board

343Hz

raw frame rate

0.39N

normal RMSE, 1–5 N

0.16N

shear RMSE, up to 1.5 N

real time

Force accuracy

Decoded forces track the F/T reference

0.43 N normal · 0.22 N shear RMSE after a 5-minute calibration

2×

Sensitivity

Feels a scalpel touch a balloon, and stops

0.1 N shear threshold · 20/20 retracted before puncture

4×

Policy learning

Shear guides a learned plug insertion

88% success with 3-DoF touch, vs 48% with vision only

Abstract

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 4×4 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. Design files, firmware, software and fabrication guide will be released upon paper acceptance.

Watch

VecTile in three minutes

01 · Sensing principle

A nub turns shear into pressure

Each silicone nub sits over four pressure-sensitive quadrants. Pressing loads all four equally; shear tilts the nub and loads one side more.

Electronics sense normal pressure. A resistive fabric on a flexible PCB reads 64 pressure quadrants, scanned by an Arduino Nano at 343 Hz.
Passive silicone nubs turn shear into pressure. Each nub sits over four quadrants: their sum gives normal force and the difference between opposite sides gives shear, decoded to [F_x, F_y, F_z] for each taxel.

VecTile design: readout board and 4x4 FPC with 2x2 quadrant cells addressed as an 8x8 matrix, transimpedance front end, and the assembled stack of FPC, resistive layer and silicone with side and top views under normal and shear force.

VecTile design. (A) The FPC implements a 4×4 taxel array with four quadrants under each nub, read as an 8×8 matrix. (B) The assembled stack: normal force raises the four quadrants uniformly, while shear creates a difference across them.

02 · Simulation

Verified in simulation first

A finite element model of the silicone nub, at four heights and five preloads, isolates the mechanics from the electronics. Normal and shear act along nearly perpendicular directions in the (Σ, D) reading space.

Decoupled

Shear and normal don't get mixed up. Pressing moves the sum Σ, shearing moves the difference D, and the two respond at 86–90° to each other.

Solvable

Forces can be recovered from the readings. Every simulated load maps to its own (Σ, D) reading, so normal and shear force can be solved back out.

Taller is better

A taller nub gives a stronger shear signal. 1 N of shear on the 3.0 mm nub shifts 1.47 N between its halves, while 89% of the normal force still reaches the sensor.

Contact pressure under the nub during the shear ramp, for four heights × five preloads. Taller nubs push more of the load to the leading half.

Finite element analysis: loading case and pressure maps for a 3 mm nub, separability grid in (Sigma, D) space, and stiffness, normal transmission and shear gain for four nub heights.

Finite element analysis of the nub-lever mechanism. (A) Force encoding and mechanical separability for a 3.0 mm nub. (B) Effect of nub height on stiffness, normal transmission and shear gain.

03 · Fabrication

Made in-house, under $35

Only the two circuit boards are ordered: the flexible sensing PCB and the readout PCB, both from a standard PCB service. Everything else is made in-house: silicone casting in a 3D-printed PLA mold, a few glue dots and simple assembly. Besides a 3D printer, the only equipment is a hobby-grade vacuum chamber, plus a syringe, clamps and a lab scale. Every layer is firmly bonded: glue dots fix the resistive fabric to the PCB and injected silicone fuses it to the cast skin, so the stack holds together through contact-rich use.

How the mold works

Cast around a dummy. A 3D-printed dummy FPC sits in the mold while the silicone cures for 16 h, reserving a thin pocket inside the silicone layer.
Slide the sensor in. After demolding and removing the dummy, the real FPC and resistive layer slide into that pocket.
Aligned by the mold. The pocket holds the stack in place, so every nub lands on its four quadrants without manual alignment.

In the lab

The full process: mix and degas, cast and cure, glue the resistive layer, slide the stack into the silicone, inject and cure, then connect to the readout board.

Eight fabrication steps: prepare the mold set, mix and degas, layer-up and pour, cure and demold, glue microdots, lay resistive layer, slide sensor in, assemble and cure.

VecTile fabrication. A 0.4 mm printed dummy FPC reserves a pocket in the cast silicone. The real sensing stack then slides in, so the mold itself aligns every nub with its four quadrants.

	Item	USD
Sensor	FPC (30-pc order)	0.89
	Resistive layer (≈ 0.004 m²)	0.12
	Dragon Skin 30 (≈ 15 g)	0.70
	FFC cable + connector set	1.50
	Mold filament (≈ 30 g)	0.36
	Subtotal	3.57
Readout	PCB, assembled (10-pc order)	6.10
	Arduino Nano^a	25.00
	Subtotal	31.10
	Total per unit	34.67

Bill of materials. ^aA Nano-compatible ATmega328P board costs about 5 USD, bringing the total to 14.67 USD.

04 · Characterization

Taller nubs, better shear

A UR5e with an ATI Mini40 force/torque sensor presses and shears every nub of four sensors with different nub heights. Each sensor gets 16 nubs × 5 preloads × 8 shear directions, 640 contact cycles in all. The 3.0 mm sensor is also tested on convex and concave mounts.

Data collection setup with UR5e, F/T sensor, 5 mm probe and VecTile on flat, convex and concave mounts, and the five-step collection procedure repeated for 8 shear directions, 5 force levels and 16 nubs across four nub heights.

Setup and procedure. Flat and R = 85 mm curved mounts.

Data collection on the flat and curved mounts (20× speed).

RMSE (N)	Flat: nub height (mm)				Curved, 3.0 mm
	0.6	1.0	2.0	3.0	convex	concave
F_z	0.45	0.45	0.41	0.39	0.45	0.42
Shear	0.29	0.28	0.24	0.16	0.21	0.23

Held-out contacts over F_z ∈ [1, 5] N and shear up to 1.5 N. Curved results reuse the decoder fitted on the flat sensor, with no retraining: on average +0.05 N normal and +0.06 N shear.

A five-minute manual calibration

No robot needed: with VecTile on the F/T sensor, an operator presses each nub ten times with a quick press-and-circle motion. That is enough to fit the decoder.

0.43 N

normal RMSE per taxel

0.22 N

shear RMSE per taxel

With palm and fingers pressing many nubs at once (up to 15.5 N normal and 5.8 N shear), summed per-taxel predictions reach 0.71 N normal and 0.43 N shear RMSE.

05 · Applications

Shear on a robot gripper

Because VecTile bends, it mounts on a rigid fingertip or wraps onto a compliant Fin Ray tip for a Robotiq 2F-85 gripper. All of our robot experiments in this paper use the Fin Ray tips, printed in TPU, with one VecTile (3.0 mm nubs) on each finger, calibrated with the five-minute manual routine.

Stopping a scalpel before it pops a balloon

The gripper lowers a scalpel onto an inflated balloon at 5 mm/s. A per-taxel shear threshold of 0.1 N triggers retraction the moment the blade touches.

20 / 20

shear reflex: retracted before puncture

burst

normal-only threshold, same motion

Visuotactile plug insertion

A Diffusion Policy inserts a UK 3-prong plug from a wrist camera, a scene camera and both fingertip sensors. It was trained on 84 teleoperated demonstrations and tested on 25 rollouts at unseen socket positions. Shear from the pins touching the socket tells the policy which way to correct.

One rollout with full 3-DoF tactile input, in real time. Top: scene and wrist cameras and both fingertip sensors. Bottom: normal and shear force on each finger through grasp, retry, descent, seating and release.

Shear makes the difference

All 25 autonomous rollouts of each policy, played together (10× speed). Green: seated. Red: not seated.

+ 3-DoF tactile

22/25

+ normal tactile

16/25

Vision only

12/25

Successful insertions. Most of the gain is in recovering from a missed first attempt: 86% of corrections succeed with 3-DoF tactile input, versus 61% with normal-only and 35% with vision only.

Open source

Build your own, after acceptance

The design files, firmware, software and fabrication guide will be released upon paper acceptance.

HardwareSensing FPC, readout board and bill of materials.Upon acceptance

Molds & nubs3D-printable mold set for the nub heights.Upon acceptance

Firmware & softwareReadout firmware, live viewer and calibration.Upon acceptance

Fabrication guideThe step-by-step build shown above.Upon acceptance

BibTeX

@article{hou2026vectile,
  title   = {VecTile: A Low-Cost Flexible Tactile Array for 3-DoF Force Sensing},
  author  = {Hou, Jian and Zhou, Xin and Spiers, Adam J.},
  journal = {arXiv preprint arXiv:2610.31472},
  year    = {2026}
}