Force, Torque & Tactile Sensing
See also (Tier 3 family index): Sensor Families
1. At a glance
Pose-and-motion sensing ([[Robotics/sensors-pose-motion]]) tells a robot where it is. Force, torque, and tactile sensing tell it what it is touching, how hard, and where on its body. Without this channel, every contact-rich task — assembly, polishing, surgery, gripping a strawberry, handing a part to a human — devolves into open-loop hope.
Three sensor families dominate, each occupying a different point in the robot’s body:
- 6-axis force/torque (F/T) sensors at the wrist. A monolithic transducer sandwiched between the tool flange and the last joint output, measuring three orthogonal forces (F_x, F_y, F_z) and three orthogonal torques (τ_x, τ_y, τ_z) referred to the sensor’s own coordinate frame. The default measurement for any tool-in-hand contact: peg-in-hole insertion, deburring, glue-bead dispensing, surgical instrument force feedback, force-feedback teleoperation.
- Joint-torque sensors inside each joint. A strain-gauged flexure between the harmonic-drive output and the link, reading the actual torque transmitted across the joint. Enables impedance / admittance control without a wrist sensor (
[[Robotics/impedance-control]]), exposes external contact anywhere on the arm (not just the tool), and is the basis for sensitive collision detection in cobots. Pioneered by DLR LWR (Albu-Schäffer 2007) and now standard on KUKA iiwa and Franka Panda. - Tactile / skin sensors on the gripper or body. Spatially distributed pressure measurement: an M×N grid of capacitive, piezoresistive, magnetic, or optical sensing cells embedded in a soft surface. Reports the pressure distribution and (in multi-axis variants) shear, vibration, and thermal flux. Used for slip detection, in-hand object localisation, grasp-quality estimation, dexterous manipulation, and safety skins for human-robot interaction.
A fourth, narrower category — single-axis force sensors (load cells, button cells, beam transducers) — handles drawer-pull tests, end-effector tip force, and prosthetic finger-tip force; mechanically simpler but identical physics.
First ask before specifying any of these:
- Where is the contact, and what frame do you want the measurement in? Wrist sensors miss elbow contact; joint torque sensors expose it.
- What is the maximum impact load, not the steady-state load? Saturation during a 10 ms impulse breaks more sensors than years of nominal duty.
- What is the smallest force you must render or measure? Sets noise-floor, drives ADC bits and signal-conditioning budget — typically 1/10 of the smallest force.
- What is the bandwidth? Rigid contact has spectral content to 1–2 kHz; the sensor and its acquisition chain must match.
2. First principles
Strain-gauge transduction
The workhorse of every wrist F/T sensor, every joint-torque sensor, and most beam load cells. A strain gauge is a serpentine resistive grid bonded to a structural element; when the element deforms, the grid’s length and cross-section change, and so does its resistance:
R = R₀ · (1 + GF · ε)
where ε = ΔL/L is the strain, R₀ is the unstrained resistance (350 Ω is the industrial standard, 120 Ω the legacy default, 1 kΩ for low-current designs), and GF is the gauge factor:
- Foil (constantan) gauges: GF ≈ 2.0, linear, low TCR, mature manufacturing. Window Industries, Vishay Micro-Measurements, HBM, Kyowa.
- Semiconductor (p-type silicon) gauges: GF ≈ 50–200, factor 25–100× sensitivity at the cost of strong temperature dependence and nonlinearity. Used where signal-to-noise must beat thermal noise of the bridge resistors (sub-N forces, MEMS pressure sensors).
- Thin-film (sputtered) gauges: GF ≈ 2, fabricated directly onto the steel flexure by sputter deposition. Lower hysteresis than glued foil; standard on premium F/T sensors (ATI Nano17, Bota SensONE).
Gauges are wired in a Wheatstone bridge to produce a differential output proportional to strain difference and to first-order cancel temperature drift, supply variation, and common-mode bending:
V_out = V_excit · (GF · ε / 4) · k_bridge
where k_bridge is 1 for a quarter-bridge (one active gauge), 2 for a half-bridge (two active gauges in opposite arms), and 4 for a full bridge (four active gauges). A full bridge at GF = 2, ε = 1000 µε (a typical full-scale strain), V_excit = 5 V produces V_out = 5 · 2 · 0.001 · 4 / 4 = 10 mV — the familiar 2 mV/V full-scale rating. See [[Engineering/op-amps]] for instrumentation-amplifier conditioning.
Bridge configurations. Four configurations recur in F/T and joint-torque sensors:
- Quarter-bridge (one active gauge + three fixed reference resistors): cheapest, most temperature-sensitive (resistor temperature coefficient does not cancel). Used in low-cost single-axis load cells.
- Half-bridge (two active gauges in opposite arms on opposite sides of the strained element): two-fold sensitivity, first-order temperature compensation. Standard in beam-style load cells.
- Full bridge with two tension + two compression gauges: four-fold sensitivity vs single gauge, full first-order temperature compensation, rejects axial load coupling. Standard in 6-axis F/T sensor arms and joint-torque tubes.
- Multi-bridge with shear-pattern gauges at 45°: maximises shear-strain sensitivity for torque measurement. Used in cross-flexure joint-torque sensors.
Capacitive tactile transduction
A parallel-plate capacitor has C = ε · A / d. Press on the upper plate, the dielectric (air gap or elastomer) compresses, d decreases, C increases:
ΔC / C₀ = − Δd / d₀ ≈ (k_elastomer / d₀) · F
for a soft dielectric of stiffness k_elastomer under axial force F. Capacitance changes of 1–10 pF over a quiescent 1–100 pF are typical. Multi-axis variants use multiple electrode pads under one elastomer dome and decompose normal vs shear from differential capacitance. Commercial implementations: Pressure Profile Systems (PPS), Tekscan capacitive sheet, XELA uSkin (uses capacitance plus a magnet for triaxial). Read out with a capacitance-to-digital converter (AD7745, FDC2214, MAX44009-class touch-screen controllers repurposed).
Piezoresistive (silicon MEMS) transduction
A silicon membrane or beam, doped to be conductive and patterned with current-flow channels, changes resistance under strain through the piezoresistive effect in crystalline silicon — physically the same phenomenon as the semiconductor gauge factor, but lithographically defined directly on the MEMS die rather than glued onto a steel flexure. Foundational physics: shear-induced changes in conduction-band geometry; see [[Engineering/semiconductor-devices]].
Used in low-cost FSR (force-sensing resistor) sheets (Interlink, Tekscan FlexiForce — actually a polymer thick-film piezoresistor, not silicon) and in MEMS barometric/pressure sensors retasked as tactile cells (Bosch BMP series under a button).
Piezoelectric readout — charge amplifier. Because PVDF film generates charge rather than voltage, the standard front end is a charge amplifier: an op-amp in inverting configuration with a capacitor (typical 100 pF – 1 nF) in the feedback path. Output voltage V_out = − Q / C_f, independent of cable capacitance. A high-impedance feedback resistor (1–10 GΩ) sets the high-pass cutoff. The op-amp must have very low input bias current (FET-input parts: AD8606, OPA197, MAX44272) to avoid leaking the charge before it can be read. Standard in BioTac and most PVDF-based slip sensors.
Optical (camera-based) tactile transduction
A miniature camera looks at the inside of a deformable elastomer membrane. The membrane has a printed marker pattern on its inner surface or is coated with a thin reflective layer. Contact deforms the membrane; the camera images the resulting marker displacement field at 30–90 fps. Downstream computer vision (often a small CNN or photometric stereo) recovers a dense 3D shape map and from that the contact pressure distribution and shear field.
Variants:
- GelSight (Adelson MIT 2009, commercialised by GelSight Inc): photometric stereo from three coloured LEDs at different angles. Spatial resolution sub-millimetre, force resolution mN-class, frame rate 25–60 fps.
- DIGIT (Lambeta et al, Meta AI 2020): a $15 fingertip-scale GelSight variant; open-source hardware and software at digit.ml.
- DenseTact (Stanford 2022): full-hemispherical fingertip with similar photometric principles.
- TacTip (Bristol Robotics Laboratory 2017): tracks pin tips on the membrane’s interior rather than a continuous surface pattern.
- GelSight Mini / Wedge (commercial GelSight Inc product line 2022 onwards).
Optical tactile sensors deliver the highest spatial resolution of any tactile family but at the cost of bulk (camera + LEDs + lens behind the membrane), LED ageing requiring annual recalibration, and tens-of-ms latency limited by camera frame rate.
Magnetic tactile transduction
A small permanent magnet (typically a 2–3 mm NdFeB cylinder, diametrally or axially magnetised) is embedded in a soft elastomer dome. Underneath, a 3-axis Hall or AMR sensor IC (Melexis MLX90393, AKM AK09918) reads the magnet’s displacement field as the dome deforms. Pressing axially moves the magnet down (changes B_z field); shearing moves it laterally (changes B_x, B_y). Combining 4–16 such cells in a tactile array gives shear-capable skin with cell pitches of 4–8 mm. Commercial: XELA Robotics uSkin (Tomo et al 2018), ReSkin (Meta AI 2021 — uses unstructured magnetised elastomer + ML decoding rather than a discrete magnet per cell).
The magnetic approach has two attractive properties: (1) the sensor electronics are separated from the contact surface by the elastomer, so the contact layer is a cheap replaceable consumable; (2) shear sensitivity comes “for free” from the 3-axis field reading without any extra electrodes. The trade-off is sensitivity to external magnetic fields — a nearby motor or steel structure shifts the baseline, requiring magnetic shielding (mu-metal cup behind the IC) or active baseline tracking.
Capacitive readout circuits. Two common topologies:
- Single-ended switched-capacitor (FDC2214, AD7745): drives the sense electrode through a known reference capacitor and measures the resulting voltage at a fixed phase of a charge-redistribution clock. Resolution ~aF, immune to parasitic capacitances when properly shielded. Standard in modern tactile arrays.
- Differential ratio measurement: two adjacent cells share a common excitation; the ratio (C₁ − C₂) / (C₁ + C₂) is read directly. Cancels common-mode parasitic effects (temperature, humidity). Used in high-end MEMS accelerometers and some shear-capable tactile cells.
Piezoelectric transduction
PVDF (polyvinylidene fluoride) film and PZT (lead zirconate titanate) ceramic generate a voltage proportional to the rate of change of applied force, with no DC response. Excellent bandwidth (DC–MHz limited by the readout amplifier) and ruggedness, but useless for static measurement. Used as slip / vibration sensors in fingertips (Wettels/Loeb BioTac 2008), as fast contact-event triggers, and for spectrum-analysis-based slip detection (Fishel & Loeb 2012: spectral energy in 30–200 Hz band detects incipient slip before macroscopic motion).
Optical force (deformation-based, non-camera)
Distinct from the camera-based optical tactile sensors above, a small family of non-imaging optical force sensors uses a single LED + photodiode pair separated by a deformable elastomer pad. Pressing the pad compresses the elastomer, changing the LED-to-detector light coupling. Output is a single analog voltage proportional to force on one axis.
Commercial: Optoforce OMD series (now part of OnRobot), which extends the principle to 3-axis by reading three photodiodes around a hemispherical reflective dome. The OMD’s reported advantages over strain-gauge designs: massive overload tolerance (200 % rated FS without damage, sometimes more), no electrical drift over temperature (the photodiode reading is ratiometric), and very low hysteresis. Disadvantages: 3-axis only (no torque), lower bandwidth (~200 Hz), and sensitivity to ambient light leaks through worn dome surfaces. Useful where a heavy industrial arm needs simple contact force without 6-axis cost.
Six-axis F/T sensor — mechanics
A 6-axis wrist sensor is a single steel (often Aluminium 7075 or maraging steel) monolithic body machined into a Maltese-cross or Stewart-platform topology. The cross arms or struts are instrumented with multiple strain-gauge full bridges, typically 6–8 bridges, each maximally sensitive to one component of the 6-vector wrench (F_x, F_y, F_z, τ_x, τ_y, τ_z) and minimally sensitive to the other five.
Because no real flexure perfectly isolates a single axis, every sensor has finite cross-coupling. The raw bridge outputs b ∈ ℝ⁶ are mapped to the wrench w = [F_x, F_y, F_z, τ_x, τ_y, τ_z]ᵀ via a factory-calibrated decoupling matrix C ∈ ℝ⁶ˣ⁶:
w = C · b + w_offset
C is determined by applying known pure-axis loads at the factory (typically 6 loadings × 3 sign combinations) and least-squares-fitting. C ships in the sensor’s flash memory or in a calibration certificate; the host reads it at boot. Cross-coupling residuals after calibration are typically 1–2 % full-scale per axis (ATI Mini40 datasheet); precision sensors achieve < 0.5 % (Bota SensONE, ATI Nano series).
Joint-torque sensor — mechanics
A torque-tube or cross-flexure element is inserted between the harmonic-drive output and the link. Strain gauges on the deformable section read the angular twist of the element under torque load:
ε ≈ τ · r / (G · J)
where r is the gauge radius from the torque axis, G is the shear modulus, J is the polar moment of area. For a typical cobot joint (50 N·m max), the flexure is sized to produce ~1000 µε at full scale, yielding 2 mV/V into a full bridge.
Stiffness trade. Stiff flexure: low noise (small angular deflection, less coupled vibration), poor resolution (small strain at full scale). Compliant flexure: high resolution, but adds a series-elastic element that lowers the joint’s overall position bandwidth and introduces a resonance in the structural mode at ω_n = √(k_flex / J_load). Industry compromise: ~3000–5000 N·m/rad joint stiffness at the sensor, ~50× softer than a rigid steel coupling but stiff enough that resonances stay above 100 Hz.
Tactile array readout
An M × N grid of tactile cells is most often read by row-column scanning: drive one row with the excitation signal, read all N columns simultaneously through a multiplexed ADC, then advance to the next row. Total scan time = M · T_settle where T_settle (5–100 µs per cell) is dominated by RC settling for capacitive arrays and conversion time for piezoresistive arrays. A 16×16 grid at 50 µs/row settles in 800 µs → ~1 kHz frame rate per cell. Alternative: individually addressed cells with on-chip multiplexers (Tekscan tile chips, BeBop Sensors fabric).
Jacobian-transpose wrench reconstruction (joint-torque to Cartesian)
When the robot has joint-torque sensing but no wrist F/T sensor, the external wrench on the end-effector can be reconstructed from the joint-torque vector τ ∈ ℝⁿ via the manipulator Jacobian J ∈ ℝ⁶ˣⁿ:
w_ext = (J · J^T)^{-1} · J · (τ_meas − τ_grav(q) − τ_friction(q, q̇) − M(q) · q̈ − C(q, q̇) · q̇)
The subtractions remove gravity (computed from the URDF), friction (from a calibrated friction model — typically Coulomb + viscous + Stribeck), and inertial torques (computed via the recursive Newton-Euler algorithm). The remaining torque is what the environment applied; the pseudo-inverse maps it to the 6-vector wrench at the end-effector. This is the architecture used by KUKA iiwa and Franka Panda — see [[Robotics/dynamics-rigid-body]] for the dynamics terms, [[Robotics/impedance-control]] for the closed-loop architecture.
Practical caveats: (1) the Jacobian is singular at workspace boundaries and the pseudo-inverse blows up — clamp the wrench output or switch to a Levenberg-Marquardt-damped inverse; (2) friction model errors at low velocities (the Stribeck transition) inject a ~10–20 % torque error that masquerades as external wrench; (3) inertial torque dominates during fast motion and is hard to identify accurately — most cobots reduce gains during high-velocity motion and rely on contact wrench only at low speed.
3. Practical math / worked examples
Numerical sanity-check before any other math: the smallest force the sensor must resolve is at least 3–5× its 1-σ noise floor; the largest impact force is at least 2–3× the steady-state design force. These margins absorb temperature drift, calibration aging, and the unmeasured impact loads of real assembly tasks. Skipping either margin guarantees field failure within months.
Worked example A — wrist F/T sensor selection for peg-in-hole
A cobot must insert a precision aluminium peg (20 mm diameter) into a chamfered hole with a clearance of 50 µm, friction coefficient µ = 0.15, and a maximum axial assembly force of 50 N (limited by what the joint impedance loop can safely exert without buckling the part).
Worst-case lateral force during sliding contact at the hole rim:
F_xy_max = µ · F_z = 0.15 · 50 = 7.5 N
Torque about the tool z-axis from off-centre contact at the chamfer radius r = 10 mm:
τ_z_max = µ · F_z · r = 0.15 · 50 · 0.01 = 0.075 N·m
Torque about the lateral axes from the moment arm of the friction force over the peg-grip length L = 30 mm:
τ_xy_max ≈ F_xy · L = 7.5 · 0.03 = 0.225 N·m
The smallest force the controller must resolve to keep the impedance loop crisp is roughly 1/10 of the smallest commanded force — call it 0.1 N. Required force resolution: 0.1 N. Required torque resolution: similar fraction of 0.075 N·m → 0.005 N·m.
Candidate: ATI Mini40 SI-40-2 calibration. Datasheet specs: F_x, F_y ±40 N (resolution 1/50 N = 0.02 N), F_z ±120 N (resolution 1/20 N = 0.05 N), τ_x, τ_y, τ_z ±2 N·m (resolution 1/2000 N·m = 0.0005 N·m). Ample margin on every axis; the limiting factor is the F_z full-scale of 120 N against a possible 200 N impact spike — install a mechanical compression stop or use a Mini45 (±290 N axial).
Worked example B — joint-torque sensor noise budget
A 6-DOF cobot joint with maximum operating torque τ_max = 50 N·m must resolve τ_min = 0.01 N·m (the smallest meaningful gravity-compensation error on a 5 kg payload at 200 mm offset is m·g·r ≈ 10 N·m, so 0.01 N·m is 0.1 % accuracy — typical for cobots).
Sensor design: a full Wheatstone bridge of four 350 Ω foil gauges, gauge factor 2, on a steel torque tube sized to produce 1000 µε at 50 N·m.
Bridge output at full scale, excitation 5 V:
V_FS = (GF · ε_FS / 4) · k_bridge · V_excit = (2 · 0.001 / 4) · 4 · 5 = 10 mV
Bridge output at minimum-resolvable torque (0.01 N·m = τ_max / 5000):
V_min = 10 mV / 5000 = 2 µV
Conditioning: Texas Instruments ADS1262 32-bit Σ-Δ ADC with integrated PGA (programmable gain amplifier) at gain 32×.
Input-referred noise (PGA + ADC) at 100 SPS: ~ 6 nV_rms (datasheet table 7-9). Effective resolution: V_min / noise = 2 µV / 6 nV ≈ 333:1 SNR at full bandwidth → ~9 effective bits at 100 Hz update. To raise effective bits to 11–12 (matching the 4096:1 dynamic range we want), oversample to 20 SPS or apply a 50 Hz hardware LPF in front of the PGA.
Thermal-EMF gotcha. Junctions between copper and the steel torque-tube material (typically a Kovar or invar steel) generate a thermal-EMF of ~5 µV/°C. A 1 °C gradient across the bridge corrupts the reading by 0.5 µV — equivalent to ~0.25 % full-scale. Mitigations: matched-pair lead-out wires, low-thermal-EMF terminal blocks, auto-tare on idle, and a continuous temperature-compensation table indexed by the bridge-substrate thermistor reading.
Worked example B-supplement — joint-torque sensor stiffness and resonance
Continuing the cobot joint above (50 N·m max, full bridge sized to 1000 µε FS). Pick a torque-tube outer radius r = 25 mm, wall thickness t = 3 mm, and material steel (G = 80 GPa). Polar moment of area:
J ≈ 2π · r³ · t = 2π · (0.025)³ · 0.003 ≈ 2.95 × 10⁻⁷ m⁴
Torsional stiffness over a tube length L = 30 mm:
k = G · J / L = 80 × 10⁹ · 2.95 × 10⁻⁷ / 0.03 ≈ 7.85 × 10⁵ N·m/rad
Strain at the outer surface at full-scale torque:
ε = τ · r / (G · J) = 50 · 0.025 / (80 × 10⁹ · 2.95 × 10⁻⁷) ≈ 5.3 × 10⁻⁶ rad on the surface
Too low — strain is below 10 µε, the bridge output would be only ~50 µV at full scale, swamped by noise. Reduce wall thickness to 1 mm and length to 40 mm: J ≈ 9.8 × 10⁻⁸ m⁴, k ≈ 1.96 × 10⁵ N·m/rad, surface strain at full FS ≈ 1.6 × 10⁻⁵ rad. Still tight. Use shear-strain gauges at 45° aligned to the principal strains, which doubles output, and accept that production designs use complex flexure geometries (cross-roller, web, cup-cantilever) rather than a simple tube. KUKA iiwa’s transducer is a webbed cross with k ≈ 16 000 N·m/rad and ~1000 µε FS — far softer than this tube example because joint stiffness, not strain sensitivity, was the design driver.
The link inertia J_link ≈ 0.5 kg·m² for the iiwa elbow gives a structural mode at:
f_n = (1 / 2π) · √(k / J_link) = (1 / 2π) · √(16000 / 0.5) ≈ 28 Hz
The position loop bandwidth is then capped at ~6 Hz unless active damping is added. iiwa’s controller uses inner-loop torque control plus a structural-mode damping injection to extend the effective bandwidth to ~50 Hz.
Worked example C — tactile array bandwidth and bus
A robot fingertip carries a 16×16 capacitive tactile pad (256 cells, 3 mm pitch, 48 mm × 48 mm coverage). Each cell has a baseline capacitance of 5 pF and a press-induced ΔC up to 1 pF. The control loop wants 100 Hz pressure update.
Raw data rate at one 12-bit reading per cell:
256 cells · 100 Hz · 12 bits = 307 kbit/s
Row-column scan: 16 rows × 16 columns. With a 5 µs settle per column × 16 columns parallelised through the ADC mux, one row takes 80 µs → 16 rows = 1.28 ms per frame → 781 Hz potential frame rate, well above the 100 Hz target.
Bus selection:
- I²C @ 400 kHz Fast-Mode: carries 50 kbyte/s usable → 9.6 kbyte/s @ 100 Hz → 256 cells × 2 bytes = 512 bytes per frame × 100 Hz = 51.2 kbyte/s. Marginal.
- SPI @ 10 MHz: 1 Mbyte/s usable. Comfortable.
- USB HID @ 1 kHz polling: 1024 bytes/poll × 1 kHz = 1 Mbyte/s. Standard for desktop-tethered tactile sensors (e.g. Pressure Profile Systems “Robotouch”).
- CAN-FD @ 5 Mbps: robust over the inside of a robot arm cable harness, 500 kbyte/s usable. Common in industrial tactile fingers.
Pick SPI over CAN-FD if the sensor MCU is < 30 cm from the host; CAN-FD if the wiring runs through multiple joints. EtherCAT ([[Languages/Tier3/robotics-control]]) for whole-arm tactile-skin networks (Mittendorfer CellulARSkin 2011 prototyped on CAN; production HumanSkin-Lab on EtherCAT).
Worked example D — slip detection from a piezoelectric / accelerometer fingertip
A robot grasps a 200 g object with 5 N grip force. Coulomb friction (µ_static = 0.4) holds the object against gravity until tangential force at the contact exceeds µ_s · F_n = 2 N. As external load (e.g. a human pulling on the object) approaches the slip threshold, microscopic slip events propagate from the edge of the contact patch inward, generating broadband mechanical vibration in the 30–500 Hz band (Westling & Johansson 1984 on human fingertips; Fishel & Loeb 2012 on the BioTac).
Detection: a PVDF film or MEMS accelerometer (e.g. ADXL356) embedded in the fingertip is high-pass-filtered at 30 Hz, rectified, and integrated over a 10 ms window. Slip event = window energy exceeds 3× the no-contact noise floor. Reaction: clamp grip force by an extra 20 % on detection. Total latency from slip onset to corrective grip command: ~15–25 ms, well below the 100–200 ms required to prevent macroscopic drop.
This is the basis of the slip-suppression control loops in research grippers (BioTac-equipped Shadow Hand, Yale OpenHand with PVDF film) and is increasingly bundled with capacitive tactile arrays in commercial fingertips (Contactile PapillArray, Robotous Tactile Skin).
Worked example D-supplement — incipient slip threshold
The biology baseline: human fingertips detect incipient slip at micro-displacements of ~5 µm by sensing skin shear and high-frequency vibration via Pacinian and Meissner corpuscles (Westling & Johansson 1984). Robots target an order of magnitude worse — ~50 µm displacement — and rely on the spectral content rather than the displacement itself.
A robot fingertip with a PVDF film senses a contact onset spike spanning 0.5–2 ms; the integrated charge represents the contact impulse. Once contact is steady, the PVDF output decays to baseline (no DC response). When the held object starts to slip, the friction-induced micro-vibrations re-excite the film. Empirical threshold: integrated rectified output over a 5–10 ms window exceeds 3σ of the baseline for two consecutive windows.
Latency budget: PVDF response 1 ms + charge-amp filter 2 ms + ADC + comm 1 ms + decision 1 ms + grip-force command 3 ms + actuator response 5 ms = ~13 ms. Slip detection latency below 20 ms reliably prevents loss-of-grasp on objects up to 500 g; above 30 ms the object falls.
Worked example E — capacitive-skin proximity vs contact discrimination
A safety skin on the upper arm of a cobot (PFL category, ISO/TS 15066:2025) uses 64 capacitive cells, 50 mm × 50 mm each, on a flexible substrate. The same cell can measure (a) proximity to a grounded conductor (human hand) via self-capacitance change up to ~50 mm away, and (b) contact force via mutual-capacitance change between the cell electrode and a pressed-down upper conductor layer.
Baseline self-capacitance C_self,0 ≈ 80 pF. Approach by a human hand at 30 mm distance: ΔC ≈ 0.5 pF (∼0.6 %). Direct contact and pressing with 5 N force: ΔC ≈ 5–10 pF (∼6–12 %). The cell driver (FDC2214, ~30 aF resolution per channel) easily resolves both. The safety controller emits two output channels:
- Proximity output triggers a velocity reduction when any cell sees > 0.2 pF ΔC.
- Contact output triggers an immediate stop when any cell sees > 2 pF ΔC.
The control logic implements two-stage de-escalation: detect-and-slow at long range, then detect-and-stop at touch. This pattern is standard in modern safety-rated cobot skins (KUKA Cybertouch, Bosch APAS, Pilz SafetyEYE3D paired with a contact skin).
4. Design heuristics
Selection rules first
Before picking a part, work out which family the application needs:
- Is the contact location predictable (always at the tool)? → wrist F/T sensor; cheapest, fastest, lowest software burden.
- Could contact occur anywhere on the arm (e.g. cobot bumping a human’s torso)? → joint-torque sensors (or a model-based current estimate for collision detection only).
- Do you need to know where on the surface contact occurred and how it’s distributed? → tactile array.
- Do you need to detect slip or surface texture? → vibration sensor (piezoelectric or accelerometer) integrated into the fingertip.
- Do you need safety-rated contact detection? → a redundant pair, typically joint-torque + safety-rated skin, certified to PLd Cat 3 or higher per ISO 13849-1.
A single robot frequently has all three: a wrist F/T for tool-side feedback, joint-torque sensors at every joint for whole-arm compliance and collision detection, and tactile-array fingertips for in-hand manipulation. The KUKA iiwa + Schunk SVH hand + Bota wrist sensor + DLR tactile fingertips combination is a common research stack.
Selection — F/T vs joint-torque vs tactile
| Goal | Sensor of choice | Why |
|---|---|---|
| Tool-side force feedback for assembly | Wrist 6-axis F/T | Closest to contact; minimal kinematics-induced noise. |
| Whole-arm collision detection | Joint-torque (or current-based estimate) | Wrist sensor blind to elbow / link contact. |
| Sensitive impedance / compliance over the whole workspace | Joint-torque | Direct measurement of transmitted torque; no kinematic chain to invert. |
| Slip / grasp quality | Tactile fingertip | Spatially distributed pressure + shear + vibration. |
| Surface-following / polishing | Wrist F/T + admittance control | Tangential and normal force separation. |
| Surgical instrument force feedback | Strain-gauged tool shaft (single-axis or 3-axis) | Tip-side, sterilisable, miniature. |
| Human-robot safe contact (collaborative robot envelope) | Joint-torque + safety-rated skin | Defence-in-depth: torque detects firm contact, skin detects light touch. |
Wrist F/T integration
- Compliance vs sensitivity. A stiffer sensor (steel monolithic, no soft elements) gives higher position-loop bandwidth but lower force resolution at low loads. Most cobot sensors target 5–20 kN/mm axial stiffness — stiff enough to keep contact resonances above 100 Hz, compliant enough to resolve sub-newton forces.
- Crosstalk recalibration. The factory C matrix is good for ≥ 1 year if the sensor sees no impact above 50 % FS. Re-verify after any overload event. Several manufacturers (ATI, Bota) ship a recalibration sequence the user can run with a known reference weight.
- Temperature compensation. Every premium sensor has an on-board thermistor or RTD. Apply a manufacturer-supplied linear or polynomial correction: w_corr = w_raw − α · (T − T_cal). Without this, sensor drift at warm-up can reach 5–10 % FS over 30 minutes.
- Auto-tare on idle. Detect quiescent intervals (no commanded motion, low velocity, low acceleration) and re-zero the offset. This is the single biggest improvement in real-world accuracy and is implemented in essentially every cobot impedance controller.
- Hysteresis. Typical 0.5–2 % FS for a stiffness-optimised sensor; up to 5 % for cheap aluminium-flexure designs. Specify the maximum allowable hysteresis early — it sets the BOM.
- Coordinate frame. The sensor frame is not the tool TCP; transform via the constant adjoint of the tool-to-sensor transform: F_TCP = Ad_T^T · F_sensor. Mounting the sensor in a non-orthogonal orientation is a popular debugging headache; document the rotation in the URDF.
Frame conventions and the gravity wrench
The wrench reported by a wrist sensor is expressed in the sensor’s own coordinate frame, typically with Z aligned with the tool flange normal and X/Y in the flange plane. Three transforms are usually required:
- Sensor-to-TCP: a constant SE(3) transform encoding the tool’s geometry relative to the sensor mount.
- TCP-to-base: time-varying, from the robot’s forward kinematics.
- Gravity in base: a constant [0, 0, −g] vector.
The gravity wrench in the sensor frame depends on tool orientation and tool mass + centre of mass. At each robot pose, w_gravity = [m·g_sensor; r_com × m·g_sensor], where g_sensor is the gravity vector rotated into the sensor frame. Subtract this from the raw measurement before reporting the external contact wrench to the impedance controller. Skipping this step makes the impedance controller try to push the tool’s own weight away, producing a recognisable “the arm sags toward gravity” failure mode.
Identifying tool mass and CoM: command the robot through 6–10 random poses, record the raw wrench, and least-squares fit (m, r_com_x, r_com_y, r_com_z). Modern robot controllers (UR, Franka, KUKA) automate this as a “payload identification” wizard.
Joint-torque integration
- Series-elastic resonance. Adding a torque sensor adds a spring; the resulting system has a structural mode at ω_n = √(k_sensor / J_link). Keep ω_n ≥ 5× the desired position-loop bandwidth (KUKA iiwa: sensor stiffness ~16 000 N·m/rad, link inertias ~0.5 kg·m², ω_n ≈ 180 rad/s ≈ 28 Hz — adequate for the iiwa’s 200 Hz position bandwidth via inner-loop damping).
- Bandwidth limit. Joint-torque acquisition needs ≥ 1 kHz update for impedance control; sample with a 24-bit Σ-Δ ADC at 1–2 kSPS through a hardware anti-alias filter at 500 Hz.
- Current-based estimate as a cheap substitute. UR e-series, most consumer cobots, and many open-source arms (Open-Manipulator, MOVE robots) skip the strain gauge and compute τ̂ = K_t · I_q − D · θ̇ − model from the motor current and a friction/dynamics model. Works for collision detection at 10 % torque accuracy; insufficient for precision force control.
- Output-side vs motor-side. Strain-gauge sensors at the joint output see the full kinematic chain torque including harmonic-drive friction. Sensors at the motor side (less common) miss the harmonic-drive losses but are easier to package.
Force vs torque sensor sizing
A frequent design mistake: choosing a sensor where the force range is right but the torque range is wrong, or vice versa. Manipulation tasks generate small forces and small torques (peg-in-hole: 5 N forces, 0.05 N·m torques). Drilling and milling tasks generate large forces and large torques (deburring: 50 N normal, 1 N·m around tool axis). Long-tool grasping multiplies a small contact force into a large wrist torque (a 1 N contact at 100 mm offset = 0.1 N·m wrist torque).
Worked sizing rule: the sensor torque resolution must be better than F_min · L_max, where L_max is the longest tool moment arm. For a 0.1 N target force resolution with a 100 mm tool: τ_resolution ≤ 0.1 · 0.1 = 0.01 N·m. The Mini40’s 0.0005 N·m torque resolution leaves 20× margin; a sensor with 0.05 N·m torque resolution (typical of low-end wrist sensors) is the limiting factor and would set the effective force resolution to 0.5 N.
Impedance vs admittance control — sensor implications
Two control architectures consume the same force / tactile signals but with very different sensor demands:
- Impedance control (Hogan 1985): the robot is position-controlled with high-bandwidth low-impedance behaviour; the controller renders a virtual spring/damper between commanded and actual positions, and the resulting environmental force is measured (or estimated) and used to modulate the commanded position. Force noise is benign — it just adds slight position jitter. Sensor noise floor of 0.5 N is acceptable.
- Admittance control: the robot is force-controlled; the controller takes the measured external force and integrates a desired velocity from a virtual mass-damper model. Sensor noise integrates directly into commanded velocity, then position — a 0.5 N noise floor with a 0.1 (N·s/m)⁻¹ admittance gain produces 5 mm/s velocity jitter. Force resolution must be 10× better than for impedance control on the same task.
Cobots universally use impedance (KUKA iiwa, Franka, Doosan) because their joint-torque sensors have ~0.1 N·m noise — adequate for impedance, marginal for admittance. Heavy industrial arms with wrist F/T sensors and stiff joints use admittance, exploiting the F/T sensor’s much lower noise floor.
Tactile sensor placement (dexterous manipulation)
- Dense at the fingertips. Pad densities of 1–2 mm pitch on the distal phalange enable in-hand object localisation, slip detection, and edge tracking.
- Coarser at the palm and medial phalanges. Pitch of 5–10 mm; the palm sees lower-resolution events (object presence, gross orientation).
- Vibrational sensor in fingertip. A PVDF film or MEMS accelerometer integrated under the elastomer pad to catch incipient slip in the 50–500 Hz spectrum (Fishel & Loeb 2012).
- No tactile at the wrist. Wrist-area contact is the F/T sensor’s job; redundant skin there wastes BOM.
- Safety skin coverage. Cobots (PFL category, ISO/TS 15066:2025) increasingly add a low-resolution capacitive proximity + contact skin over the entire arm. KUKA iiwa Cybertouch-skin and Bosch APAS skin are commercial examples; resolution ~50 mm but coverage 100 %.
ISO 15066 power-and-force-limiting thresholds
ISO/TS 15066:2025 specifies maximum permissible quasi-static and transient force/pressure on each region of the human body that a cobot may contact. Selected thresholds (from Annex A):
| Body region | Max quasi-static pressure (N/cm²) | Max quasi-static force (N) | Max transient force (N) |
|---|---|---|---|
| Skull / forehead | 130 | 130 | 175 |
| Face | 110 | 65 | 90 |
| Neck (back) | 140 | 150 | 210 |
| Chest (sternum) | 120 | 140 | 210 |
| Abdomen | 110 | 110 | 160 |
| Upper arm / shoulder | 160 | 150 | 220 |
| Lower arm / wrist | 190 | 160 | 220 |
| Hand / fingers | 200 | 140 | 180 |
| Thigh / knee | 220 | 220 | 320 |
| Lower leg / shin | 220 | 130 | 170 |
The cobot’s joint-torque sensors and any safety skin must detect contact below these thresholds with sufficient margin to actuate a stop within the deceleration distance. For a typical cobot moving at 0.25 m/s with 100 ms total stopping time, the contact-detection threshold is set to ~50 % of the lowest quasi-static body-region limit, ensuring stop completion before the threshold is exceeded.
Calibration and verification procedures
- Pre-deployment dead-weight calibration. Mount the wrist sensor flange-down on a fixture, attach known reference weights (0, 1, 2, 5, 10 kg traceable to a NIST-class mass), and verify F_z linearity ≤ 0.5 % FS and zero offset ≤ 1 % FS. Repeat for F_x, F_y by rotating the fixture 90°. Most premium sensors are shipped calibrated to 0.2 % FS; field re-verification should match within ± 0.5 %.
- Cross-coupling matrix verification. Apply pure-axis loads (a hanging weight along each of six axes); the cross-coupling residuals on the other five axes should be < 1.5 % FS. Persistent cross-coupling > 3 % FS indicates either a damaged flexure or an unintended cable preload. ATI sensors include a calibration self-check sequence executed at power-up.
- Tactile skin calibration. Two stages: (1) baseline — record each cell’s quiescent output at room temperature, no contact, for 5 minutes; the per-cell mean is the per-cell zero, and the std-dev is the noise floor. (2) force-pressure mapping — press a flat-faced indenter at known force on a 5×5 grid of locations, fit a per-cell gain (typically a polynomial in voltage; ADC counts → kPa). For optical (GelSight) sensors, also calibrate the photometric-stereo lookup table by pressing a hemispherical reference and learning the LED-direction-to-surface-normal mapping.
- Joint-torque sensor verification. Apply a known reactionary torque at the joint output (a horizontal lever with a calibrated weight at a known offset) and read the sensor; repeat at multiple angles to span the full-scale range. Verify the sensor reads within 1 % FS of the applied torque. Repeat with the joint at multiple temperatures (typical operating range 5–45 °C) to characterise thermal drift.
Force-control loop design
- Bandwidth. Rigid-contact force feedback resonates at 100–200 Hz mechanical mode; loop bandwidth should be ≤ 1/5 of this → 20–40 Hz. The sensor and ADC must run ≥ 10× the loop bandwidth (≥ 200 Hz, with ≥ 1 kHz preferred to leave headroom for filter group delay).
- Noise floor. The smallest force you can stably command is ≈ 3–5× the sensor 1-σ noise. For a 0.1 N noise floor, commanded forces must remain ≥ 0.5 N.
- Gain starting points for hybrid position/force or impedance control of a cobot peg-in-hole:
- Cartesian impedance stiffness K_p = 100–500 N/m (translational), 5–20 N·m/rad (rotational).
- Damping ratio ζ = 0.7–1.0 → K_d = 2·ζ·√(K_p · m) where m is the effective endpoint inertia (typically 1–5 kg for cobots) → K_d ≈ 30–50 N·s/m.
- Integral gain only on the tangential (sliding) axes; never on the constrained axes.
Mechanical mounting and overload protection
- Maximum overload before destruction. Premium sensors specify a single-overload safety factor of 5–10× FS. Below that, the flexure deforms elastically; above, plastic deformation shifts the zero offset permanently. Design a mechanical overload stop: a hard cone or pin that engages at 1.5–2× FS deflection. ATI sensors include one internally; cheaper sensors leave it as an integrator’s responsibility.
- Bolt circle and torque. Wrist sensors mount via a precise bolt circle (typically ISO 9409-1 flanges with 4 or 6 M5/M6 bolts). Apply manufacturer-specified torque (typically 6–10 N·m for M5, 12–18 N·m for M6) with a calibrated torque wrench. Loctite 243 (medium-strength, removable) is the standard threadlocker; do not use 263 (high-strength) on a sensor flange — sensor removal will damage the bolts or the flange.
- Adapter plate stiffness. Any intermediate plate between the sensor and the tool or robot adds compliance, lowers the structural resonance, and acts as a low-pass filter on the force measurement. Minimise plate count; if unavoidable, machine plates from 6061-T6 or 7075-T6 aluminium at least 8 mm thick to maintain stiffness ≥ 10× the sensor’s intrinsic stiffness.
- Cable strain relief. Mount a small bracket on the robot last link to anchor the sensor cable so cable motion at the wrist does not directly pre-load the sensor. Route the cable upstream with intentional service slack — too tight and you get configuration-dependent zero drift, too loose and the cable snags.
- Tool-side concentricity. If the tool is asymmetric or has significant offset mass, the sensor sees a static torque proportional to gravity in any non-vertical pose. Calibrate this gravity wrench at multiple orientations during the robot’s tool-change procedure and subtract it before reporting external wrench to the impedance controller.
Signal-conditioning rules (bridge + ADC)
- Differential everything. Wheatstone bridges produce a small differential voltage on a large common-mode pedestal (half the excitation). Receive with a dedicated 24-bit Σ-Δ ADC with on-chip PGA (ADS1262/63, AD7124-8, MAX11254). See
[[Engineering/op-amps]]. - Ratiometric excitation. Tie the ADC reference to the same supply that excites the bridge; the bridge ratio cancels supply variation.
- Twisted-pair shielded cable from the sensor head to the conditioner. Shield grounded at the conditioner only.
- Anti-alias at f_s / 2.5. The Σ-Δ converter’s internal SINC³ filter is not enough at high mechanical-resonance frequencies; add an external single-pole RC.
- Auto-zero (chopper) amplifier in the front end for sub-µV resolution. ADS1262 includes one; standalone parts AD8628, OPA277 are popular.
Sensor mounting orientation and conventions
Wrist sensors mount via ISO 9409-1 flanges. Common orientations:
- Z-axis along tool normal (most common): F_z is the axial push/pull on the tool, F_x and F_y are lateral forces, τ_z is rotation around the tool axis. Used by ATI, Bota, OnRobot, Schunk by default.
- Inverted Z (some old systems): F_z points away from the robot, requiring a sign flip in the controller. Verify with a known weight at boot.
- Rotated 45° around Z: a few cobot integrators mount the sensor rotated to align its strongest axis with the most-loaded tool direction. Always document in the URDF.
URDF snippet: every wrist sensor must appear as a fixed joint between the last arm link and the tool link, with explicit <origin> rpy and xyz capturing the mount transform. A common bug: forgetting to update the <origin> after switching sensors, leading to a constant 90°-rotated wrench reported to the controller. The first verification at integration is “press down on the tool; F_z should increase positively”.
Sample rate, bandwidth, and digital interface
- Sample rate ≥ 10× control rate. A 100 Hz control loop demands ≥ 1 kHz sensor sampling for clean derivative terms and to avoid extra phase lag from the anti-alias filter. Most premium F/T sensors stream at 1–8 kHz (ATI Mini40 NetBox at 7 kHz; Bota SensONE at 1 kHz native EtherCAT, 2 kHz via direct Ethernet).
- Digital interface choice. EtherCAT (1 kHz hard-real-time, integrates directly into modern robot controllers) is the production standard. EtherNet/IP and PROFINET are common in PLC-controlled cells. RS-485 and CAN-FD for embedded systems. Analog (six BNC channels, ±10 V) is legacy but persists in lab use because it’s universally interfaceable. USB is used by hobby/lab sensors but suffers from non-deterministic latency.
- Hardware timestamping. PTP / IEEE 1588 over Ethernet (or EtherCAT distributed clocks) keeps the sensor and the controller within ≤ 1 µs of each other. Without this, software-timestamped samples on a non-real-time Linux can be tens of milliseconds late, fatally degrading any closed-loop force control.
- Latency budget. End-to-end latency from physical force change to controller-side measurement should be ≤ 1 / (10 · f_control). At 100 Hz control, ≤ 1 ms. Budget: 0.1 ms sensor analog conversion + 0.3 ms ADC + 0.2 ms serialise / network + 0.4 ms controller read = ~1 ms. Anything more and the phase lag eats stability margin.
5. Components & sourcing
6-axis force/torque sensors (real parts)
| Vendor | Part / line | Ranges (F_xy / F_z / τ_xy / τ_z) | Interface | Notes |
|---|---|---|---|---|
| ATI Industrial Automation | Mini40 / Mini45 | ±40 / ±120 / ±2 / ±2 N, N·m | EtherNet/IP, EtherCAT, RS-485, analog | Industry de-facto for cobot wrists. |
| ATI Industrial Automation | Gamma | ±130 / ±400 / ±10 / ±10 | same | Mid-range industrial. |
| ATI Industrial Automation | Delta / Theta | ±660 / ±1980 / ±60 / ±60 | same | Heavy industrial arms. |
| ATI Industrial Automation | Nano17 / Nano25 | ±17 / ±40 / ±0.2 / ±0.2 | same | Sub-inch diameter, surgical / lab. |
| SCHUNK | FT / FTN / FTM series | up to ±2500 N / ±500 N·m | EtherCAT, PROFINET | German industrial. |
| Robotous | RFT, RFT80 series | up to ±200 / ±600 / ±10 / ±10 | EtherCAT, CAN | Korean, cobot-aligned. |
| Bota Systems | Rokubi, SensONE, MiniONE | up to ±200 / ±400 / ±10 / ±10 | EtherCAT (1 kHz native) | Swiss, multi-vendor cobot driver. |
| OnRobot | HEX-E / HEX-H | ±100 / ±200 / ±6 / ±6 | proprietary, UR/cobot adapters | Plug-and-play on UR/Doosan/Techman. |
| OnRobot (Optoforce) | OMD-30 / OMD-45 | up to ±450 N, 3-axis only | analog / EtherCAT | Optical-deformation principle, 200 % overload tolerant. |
| JR3 Inc | 30 / 45 / 67 mm series | up to ±2200 / ±5500 / ±200 / ±200 | PCI / EtherCAT | Long-established US vendor. |
| HBM (HBK) | K-MCS, U10M | precision load cells & 6-axis | EtherCAT, analog | Metrology grade. |
| Aidin Robotics | AFT family | up to ±200 / ±300 / ±10 / ±10 | EtherCAT, CAN | Korean cobot ecosystem. |
Joint-torque sensors
| Vendor | Part / role | Range | Notes |
|---|---|---|---|
| ATI Industrial Automation | TIA series | 25–500 N·m | Reaction-torque transducers for joint integration. |
| Robotous | RTS torque-tube | 30–500 N·m | OEM cobot joint sensor. |
| Sensitec / Lorenz Messtechnik | DR series | 0.05–10 000 N·m | German industrial reaction torque. |
| HBM | T22, T40 | 0.1–10 000 N·m | Metrology-grade in-line. |
| Built-in: KUKA LBR iiwa 7 / 14 | Cross-flexure transducer | 176 / 320 N·m per joint | Strain gauges on a torque-tube at each joint output. 0.2 % FS precision. |
| Built-in: Franka Panda / Production | Cross-roller flexure transducer | 87 N·m (J1–J4), 12 N·m (J5–J7) | Strain gauges on a parallel-spring flexure. |
| Built-in: Doosan M-series, A-series | Strain-gauge torque sensor at output stage | 40–300 N·m by joint | Production cobot torque sensing. |
| Built-in: UR e-series | Motor-current-based torque estimate — not a true torque sensor | n/a | Model-based; adequate for collision detection at PLd. |
Tactile / skin sensors
| Vendor | Part / line | Modality | Notes |
|---|---|---|---|
| GelSight Inc | GelSight Mini, Wedge | Optical (photometric stereo) | Commercial offshoot of MIT 2009 research. |
| Meta AI / CMU | DIGIT | Optical (GelSight-style) | Open-source hardware; ~$15 BOM at digit.ml. |
| SynTouch (now defunct, IP at Tangible Research) | BioTac | Pressure + vibration + thermal | 19 electrodes + hydroacoustic + thermistor. |
| Tekscan | FlexiForce A201, A301, A401 | Single-cell piezoresistive (FSR) | $20 cell, hysteresis-limited, binary use. |
| Tekscan | I-Scan and Grip sheets | Multi-cell piezoresistive sheets | 256–2300 cells per sheet, USB readout. |
| Pressure Profile Systems (PPS) | RoboTouch, FingerTPS | Capacitive sheet | Industrial / surgical glove tactile. |
| Interlink | FSR 402, 406, 408 | Single-cell piezoresistive | Sub-$10 hobby/light industrial. |
| XELA Robotics | uSkin 4x4, 6x6 | Magnetic 3-axis per cell | Triaxial, ~4 mm pitch. |
| Meta AI (Lerrel Pinto group) | ReSkin | Magnetic + ML decoding | Open-source, $5/cell elastomer + AKM sensor. |
| Bristol Robotics Lab | TacTip | Optical (pin-tracking) | Open-source, 3D-printable. |
| Contactile Pty | PapillArray | Pillar-array, 3D force per pillar | Slip detection oriented. |
| BeBop Sensors | Smart Fabric tactile | Piezoresistive textile | Wearable, glove integration. |
| Robotous | Tactile Skin | Multi-modal cobot skin | Whole-arm collision capacitive layer. |
| Pyramid / DLR HEX-O-SKIN (research) | Hexagonal modular tiles | Multi-modal (proximity + force + accel + temp) | Modular skin tiling, CAN backbone. |
Signal-conditioning ICs
| Vendor | Part | Role | Notes |
|---|---|---|---|
| Texas Instruments | ADS1262 / ADS1263 | 32-bit Σ-Δ ADC + PGA | Best in class for bridge sensors; 6 nV/√Hz noise. |
| Analog Devices | AD7124-4 / -8 | 24-bit Σ-Δ ADC + PGA | 8-channel bridge front end. |
| Maxim (now ADI) | MAX11254 / MAX11256 | 24-bit Σ-Δ ADC + PGA | Multichannel bridge. |
| Texas Instruments | INA826, INA826S | Instrumentation amp | Bridge front-end before discrete ADC. |
| Texas Instruments | LMP90100 | 24-bit ADC + bridge excitation | Single-chip bridge. |
| Analog Devices | AD7745 | Capacitance-to-digital | Tactile cap-array readout. |
| Texas Instruments | FDC2214 | 4-channel cap-to-digital | Capacitive tactile / proximity. |
Single-axis load cells (for completeness)
Many simpler robotics tasks need only one force axis: a button-press tester, a drawer-pull rig, a strain-gauged surgical instrument shaft, a prosthetic-fingertip force read-out. The single-axis market spans:
| Form factor | Range | Vendors | Typical price |
|---|---|---|---|
| Button / load button | 0.1 N – 50 kN | Honeywell Model 13, Interface SM, Futek LLB | $50–500 |
| S-beam | 0.5 N – 100 kN | HBM U2A, Tedea-Huntleigh, Futek LSB | $100–800 |
| Beam (bending) | 0.5 N – 5 kN | Vishay 1004, HBM PW6D | $40–300 |
| Pancake / low-profile | 50 N – 500 kN | Interface 1010, Futek LCM | $300–2000 |
| Tension link | 100 N – 500 kN | Crosby, Dillon | $500–5000 |
| Miniature surgical / probe | 0.01 N – 50 N | ATI Nano17 (single-axis variant), Futek LRF | $500–2000 |
These follow the same Wheatstone-bridge + ADS1262-class front end as multi-axis sensors; calibration is one constant rather than a 6×6 matrix.
6. Reference data
F/T sensor brand vs range vs resolution vs price tier
| Sensor | Diameter | F_x,y FS | F_z FS | τ_z FS | Force res. | Torque res. | Price tier |
|---|---|---|---|---|---|---|---|
| ATI Nano17 | 17 mm | ±50 N | ±70 N | ±0.5 N·m | 1/160 N | 1/24000 N·m | $$ ($3–6 k) |
| ATI Mini40 | 40 mm | ±40 N | ±120 N | ±2 N·m | 1/50 N | 1/2000 N·m | $$ ($4–7 k) |
| ATI Mini45 | 45 mm | ±145 N | ±290 N | ±5 N·m | 1/16 N | 1/750 N·m | $$ ($5–8 k) |
| ATI Gamma | 75 mm | ±130 N | ±400 N | ±10 N·m | 1/40 N | 1/1500 N·m | $$$ ($8–12 k) |
| ATI Delta | 94 mm | ±660 N | ±1980 N | ±60 N·m | 1/8 N | 1/300 N·m | $$$ ($10–15 k) |
| Bota SensONE | 70 mm | ±200 N | ±400 N | ±10 N·m | 1/80 N | 1/2000 N·m | $$$ ($8–14 k) |
| Bota Rokubi | 60 mm | ±170 N | ±340 N | ±5 N·m | 1/40 N | 1/1000 N·m | $$ ($5–9 k) |
| OnRobot HEX-E | 71 mm | ±100 N | ±200 N | ±6 N·m | 1/50 N | 1/1000 N·m | $$ ($5–7 k) |
| SCHUNK FT-Axia80 | 80 mm | ±200 N | ±400 N | ±10 N·m | 1/40 N | 1/1000 N·m | $$$ ($8–12 k) |
| Robotous RFT80-6A01 | 80 mm | ±100 N | ±200 N | ±10 N·m | 1/40 N | 1/1000 N·m | $$ ($4–7 k) |
Joint-torque sensors in modern cobots
| Robot | Per-joint torque sensor? | Sensor type | Stiffness, FS torque |
|---|---|---|---|
| KUKA LBR iiwa 7 R800 / 14 R820 | Yes, all 7 joints | Cross-flexure strain-gauge transducer | ~16 000 N·m/rad, 176/320 N·m |
| Franka Production / Research | Yes, all 7 joints | Cross-roller flexure strain gauge | ~10 000 N·m/rad, 87 N·m (J1–4) |
| Doosan M0609 / A0509 | Yes, all 6 joints | Strain-gauge torque sensor | ~12 000 N·m/rad, 60–200 N·m |
| Kinova Gen3 | Yes, all 7 joints | Custom strain-gauge transducer | ~8 000 N·m/rad, 39 N·m |
| Rethink Sawyer (discontinued) | Yes, all 7 joints | Series-elastic actuator | Compliant by design, 80 N·m |
| Universal Robots e-series (UR3e, UR5e, UR10e, UR16e, UR20) | No — model-based current estimate | n/a | n/a |
| Techman TM5, TM12 | No — model-based estimate | n/a | n/a |
| Fanuc CR-7iA, CRX-10iA | Optional add-on F/T sensor at wrist | n/a (no per-joint) | n/a |
| Universal Robots e-series with Bota / OnRobot HEX wrist sensor | Wrist only, no per-joint | 6-axis F/T at flange | sensor-dependent |
Tactile sensor technology comparison
| Family | Spatial res | Bandwidth | Static? | Shear? | Robustness | Cost |
|---|---|---|---|---|---|---|
| Capacitive sheet (PPS, Tekscan cap) | 1–5 mm | 100 Hz–1 kHz | Yes | No (single-axis variant) | Moderate (membrane wear) | $$ |
| Capacitive shear (XELA, sometimes via magnet) | 3–8 mm | 100 Hz–1 kHz | Yes | Yes | Moderate | $$$ |
| Piezoresistive FSR sheet (Tekscan, Interlink) | 2–10 mm | 100 Hz | Yes (low accuracy) | No | Low (hysteresis 5–20 %) | $ |
| Piezoresistive MEMS array | 0.5–2 mm | 1 kHz | Yes | No | Moderate | $$ |
| Optical / GelSight | 0.05–0.5 mm | 30–60 Hz | Yes | Yes (full 3D field) | Low (gel wears) | $$ ($15–500 / unit) |
| Magnetic (XELA uSkin, ReSkin) | 4–8 mm | 100 Hz–1 kHz | Yes | Yes | High | $$ |
| Piezoelectric (PVDF, PZT) | 1–5 mm | 1 Hz – 10 kHz | No (AC only) | No | High | $ |
| Multi-modal (BioTac, HEX-O-SKIN) | 1–5 mm | 30 Hz – 1 kHz | Yes | Yes + thermal + vibration | Moderate | $$$$ |
Slip-detection vibration channels — typical hardware
| Approach | Sensor | Bandwidth | Threshold | Notes |
|---|---|---|---|---|
| PVDF film at fingertip | Polytec PVDF or DT1-028K | 0.1 Hz – 100 kHz (charge amp) | Spectral energy 30–500 Hz | Fishel/Loeb BioTac and many research grippers. |
| MEMS accelerometer | ADXL356 (±10 g, low-noise), LSM6DSV16X | DC – 1.5 kHz | RMS energy 50–500 Hz | Used in DIGIT, GelSight Mini, recent commercial. |
| Piezoelectric stack (PZT) | Murata, Physik Instrumente | 0.1 Hz – 10 kHz | Peak pulse > 3σ | Tougher, more sensitive, larger. |
| Electret microphone | Knowles SiSonic | 20 Hz – 20 kHz | Acoustic emissions at contact onset | Rarely used; sensitive to ambient noise. |
Common signal-conditioning chips for bridge sensors
| Chip | Bits | Channels | Noise (PGA 32×) | Excitation | Notes |
|---|---|---|---|---|---|
| TI ADS1262 | 32 | 11 differential | 6 nV_rms/√Hz | external | Best-in-class load-cell front end. |
| TI ADS1263 | 32 + 24 aux | 11 + 5 | 6 nV_rms/√Hz | external | Like 1262 with auxiliary ADC for temp. |
| ADI AD7124-8 | 24 | 8 differential | 8 nV_rms/√Hz | integrated 200 µA–1 mA | Industrial bridge front-end. |
| ADI AD7794 / AD7795 | 24 | 6/4 | 21 nV_rms/√Hz | integrated 10 µA–1 mA | Legacy bridge ADC, still popular. |
| Maxim MAX11254 / MAX11256 | 24 | 6 | 12 nV_rms/√Hz | external | Cost-sensitive bridge front-end. |
| TI LMP90100 | 24 | 7 differential | 26 nV_rms/√Hz | integrated | Single-chip bridge sensor solution. |
| ADI AD8237 (instrumentation amp) | analog | 1 | 1.5 µV_rms | external | Drop-in before any 16-bit SAR ADC. |
| TI INA826 (in-amp) | analog | 1 | 0.5 µV_rms | external | High-CMRR (96 dB at G=10) bridge front. |
Force / torque resolution vs application
| Application | Force resolution | Torque resolution | Bandwidth | Typical sensor |
|---|---|---|---|---|
| Cobot peg-in-hole assembly | 0.1 N | 5 mN·m | 200 Hz | ATI Mini40, Bota SensONE |
| Surgical instrument (laparoscopic) | 0.05 N tip | n/a | 100 Hz | Strain gauge on instrument shaft |
| Polishing / deburring | 0.5 N normal, 2 N tangential | — | 30 Hz | ATI Gamma, Schunk FT |
| Industrial pick-and-place | 5 N | — | 100 Hz | OnRobot HEX, basic Robotous |
| Humanoid biped force | 5 N foot, 0.5 N·m ankle | 0.5 N·m | 200 Hz | Custom 6-axis under each foot |
| In-hand manipulation tactile | 10–50 mN per cell | n/a | 100 Hz | GelSight Mini, XELA uSkin |
| Slip detection vibration | n/a (dynamic only) | n/a | 30 Hz – 1 kHz | PVDF film, MEMS accel |
| Safety-rated PFL (ISO 15066) | 5 N | n/a | 50 Hz min | Joint torque + safety skin |
| Aerospace actuator load | 0.1 % FS | 0.1 % FS | 1 kHz | HBM K-MCS, redundant |
Tactile-array packaging trade-offs
| Choice | Pros | Cons |
|---|---|---|
| Rigid PCB with surface-mount cells | Cheap, easy to read out, mature | Cannot conform to curved surfaces; fragile in service |
| Flexible PCB (polyimide) | Conforms to curved phalanges; replaceable | More expensive; routing density limited |
| Elastomer with embedded cells (PPS, Tekscan) | Compliant; protects against impact | Replacement is messy; difficult to repair |
| Optical (GelSight) | Highest spatial resolution; full 3D field | Bulky; requires camera and LEDs; gel wears |
| Magnetic (XELA, ReSkin) | Robust; replaceable elastomer | Cell pitch limited by magnet size; needs shielding |
| Textile (BeBop fabric) | Conforms to organic shapes; washable | Low resolution; piezoresistive hysteresis |
7. Failure modes & debugging
- Slow drift over hours from temperature. Every uncompensated bridge sensor drifts 1–5 % FS over a 30-minute warm-up and a few % over a 24 h diurnal cycle. Mitigation: enable the sensor’s internal temperature compensation (every premium sensor has it; many cheap ones do not), and execute an auto-tare on idle in the controller whenever quiescent (||v|| < ε, ||a|| < ε) for ≥ 1 s.
- Saturation during impact. A drop / strike can transiently apply 10× the rated load; the strain gauges briefly see plastic deformation in the flexure, after which the zero offset shifts permanently or the sensor delaminates. Always design a mechanical overload stop: a hard cone or pin that engages at 1.5–2× FS deflection, transferring further force into a parallel stiff path. ATI and Bota datasheets specify the maximum overload before mechanical damage; observe it.
- Cable-induced parasitic torques. A stiff power/comms cable running across the wrist sensor pre-loads it. Manifests as a configuration-dependent offset that shifts when the arm moves. Mitigation: route cables with intentional slack and a pivot point upstream of the sensor; tare in each working pose, or model the cable-induced offset as a function of joint angle and subtract.
- Mounting backlash / poor bolt preload. Bolts torqued below spec let the sensor flexure micro-shift under load → 1–5 % hysteresis. Apply manufacturer-specified bolt torque (typically 15–25 N·m for M5–M6 mounting bolts on a Mini40-class sensor), use Loctite 243, and verify with a single load-unload cycle that the residual offset is < 0.5 % FS.
- Strain-gauge depoling / fatigue. Repeated overload above 70 % FS accelerates fatigue failure (Coffin-Manson). After 10⁶ cycles, expect drift. Replace the sensor when calibration C matrix residuals exceed 5 %.
- Bridge resistor mismatch from solder reflow stress. During PCB assembly the strain gauges are sometimes near reflow zones; thermal stress during solder cures changes bridge balance and produces a zero offset that does not respond to taring. Verify bridge balance on a manufactured sensor before final encapsulation.
- Salt-water / chemical exposure. Foil gauges and their bonding adhesive degrade rapidly in salt water, acidic environments, or hydraulic-fluid spills. Sensors for marine or process-industry use need explicit chemical-resistance specs (IP67, IP69K, hermetic seal). The Bota Rokubi-W is IP67-rated; ATI offers a “Wet” variant.
- Resonant amplification of motor PWM. A motor running at a frequency near the sensor flexure’s structural mode pumps energy into the sensor; the bridge output shows oscillations at the resonance frequency, not the PWM frequency. Mitigation: shift the PWM frequency away from any flexure mode (the Bota datasheet lists each sensor’s first three modes); add mechanical damping at the mount.
- EMI / PWM coupling on the bridge. Motor PWM at the wrist generates dV/dt-coupled glitches on the bridge cable; manifests as a low-frequency offset that scales with motor duty cycle. Mitigations: twisted-pair shielded cable, common-mode choke at the receiver, ferrite bead at the connector, LPF before the ADC, and physical separation from motor phase cables.
- Aliasing of structural vibration. A 200 Hz robot-arm resonance aliased by a 100 Hz sample loop appears as a few-Hz beat in the force trace. Mitigation: anti-alias at f_s/2 (RC or active LPF) and sample at ≥ 1 kHz even if the loop runs at 100 Hz.
- Coordinate-frame mismatch. The sensor’s mechanical frame is not the tool TCP; the controller computes wrenches in the wrong frame and the impedance loop oscillates in unexpected directions. Mitigation: encode the sensor-to-TCP transform in the URDF; verify by applying a known weight at the TCP and checking the predicted vs reported wrench.
- Joint-torque sensor vibration coupling. The torque-tube flexure has a structural mode at √(k_flex / J_link). Excitations of the link (impact, contact, joint reversal) ring this mode and corrupt the torque reading for the ring-down duration (~50–200 ms). Mitigation: notch filter at the resonance, or model the vibration as a separate state in a Kalman filter.
- Joint-torque sensor zero shift on power cycle. Some strain-gauge bridges have a small offset that depends on the supply voltage at startup; manifests as a 1–2 % FS zero shift between sessions. Always tare at boot, with the robot in a known unloaded pose (gravity vector known, payload subtracted).
- F/T sensor static-electricity damage. Wrist sensors mounted at the end of a robot arm with a polymer-tip tool can accumulate static charge from contact with non-conductive workpieces. Discharge through the bridge sometimes damages the strain-gauge connections. Mitigation: ESD-discharge strap to robot ground; conductive tool tips where possible.
- Tactile-array dropout cell. A single dead cell in an array biases the centre-of-pressure estimate. Detect during a baseline scan with no contact (all cells should report quiescent ± noise floor); flag and median-fill any cell that consistently reports zero or saturated.
- Tactile elastomer abrasion. GelSight and other elastomer-based skins wear after 10³–10⁴ contact events. Plan replacement as a consumable; calibrate against a flat reference after each replacement. GelSight Mini replacement gel pads are sold separately.
- GelSight LED ageing. LED output drops 10–30 % over 50 000 hours; the photometric-stereo decoder assumes constant illumination. Recalibrate annually or whenever the decoder’s baseline-image residual grows above threshold.
- Hysteresis of low-cost FSRs. Tekscan FlexiForce and Interlink FSRs have 5–20 % hysteresis and are best treated as threshold-detection elements, not analog force sensors. For analog use, calibrate every individual cell, and accept the residual hysteresis as an irreducible error.
- Magnet eccentricity in magnetic tactile cells. XELA / ReSkin cells need precise magnet centring in the elastomer dome; off-centre magnets give shear-direction bias. Recalibrate per cell after assembly; some commercial parts ship pre-characterised.
Calibration drift over service life
Premium F/T sensors hold their factory calibration to within 1 % FS for 1–3 years under normal use; cheaper sensors drift 2–5 % FS over the same period. Drivers of drift:
- Thermal cycling of the flexure — repeated expansion and contraction at the gauge bond points creates micro-cracks that change gauge resistance slowly. Manifests as a slow zero shift and a tilted decoupling matrix.
- Cumulative cyclic loading — fatigue damage near the bond. Catastrophic above 70 % FS repeated loading; tolerable below 40 % FS for ≥ 10⁷ cycles.
- Adhesive ageing — the epoxy bonding the gauges to the flexure undergoes slow creep, especially at elevated temperatures (above 60 °C). High-end sensors use sputtered gauges to bypass this; foil-gauge sensors degrade faster.
- Cable harness micro-fractures — individual conductors break inside the cable jacket after thousands of motion cycles in the robot’s cable management. Manifests as intermittent dropout on one bridge channel.
Plan re-verification annually for production cells, every 6 months for high-utilisation safety-critical applications. ATI, SCHUNK, and HBM offer factory recalibration services for $500–2000 per sensor.
Debugging recipe table
| Symptom | First check | Second check |
|---|---|---|
| F/T sensor drifts during warm-up | Temperature comp enabled? | Auto-tare on idle implemented? |
| Force reading shifts with arm pose | Cable pre-load across sensor | Joint-angle-dependent calibration table |
| Impedance loop oscillates at high gain | Sensor bandwidth too low | Frame mismatch between sensor and TCP |
| Joint-torque sensor “rings” after impact | Resonance — add notch filter | Mechanical stop saturating gauge |
| Tactile array has stuck cells | Power-rail short on row driver | Failed dome on the affected cells |
| GelSight image dark / banded | LED current dropped | Gel pad worn or detached |
| Capacitive skin reading noisy | Excitation amplitude too low | Cable capacitance shifting baseline |
| FSR reading drifts over 1 minute | Polymer viscoelastic creep | Apply pre-load and rezero |
Quick decision tree
Need contact-rich behaviour on a robot.
├── Tool-side, predictable contact location?
│ ├── Yes → wrist 6-axis F/T sensor.
│ │ ├── Light load, sub-N resolution? → ATI Nano17 / Nano25.
│ │ ├── Cobot peg-in-hole, ±100 N? → ATI Mini40 / Bota SensONE / OnRobot HEX.
│ │ ├── Polishing, ±500 N? → ATI Gamma / Schunk FT-Axia.
│ │ └── Industrial heavy, ±2000 N? → ATI Delta / Theta.
│ └── No → joint-torque or tactile.
├── Whole-arm contact, low-cost?
│ ├── Cobot with motor-current torque estimate (UR, Techman). Adequate for collision detection only.
│ └── True joint-torque sensors (KUKA iiwa, Franka, Doosan, Kinova). Required for sensitive compliance.
└── Distributed pressure / shear / slip?
├── Optical fingertip → GelSight Mini / DIGIT (research-friendly).
├── Capacitive / piezoresistive grid → PPS RoboTouch / Tekscan / XELA uSkin (commercial).
├── Magnetic 3-axis → XELA uSkin, ReSkin.
├── Slip detection only → PVDF film + charge amp, or MEMS accelerometer.
└── Safety skin for human contact → KUKA Cybertouch / Bosch APAS / custom.8. Case studies
KUKA LBR iiwa 7 R800 joint-torque sensing
The iiwa (intelligent industrial work assistant) is the production version of the DLR-LWR-III lightweight arm (Albu-Schäffer et al, IROS 2007). Each of its seven joints contains a cross-flexure strain-gauge transducer placed downstream of the harmonic-drive output, immediately before the link interface. The transducer is a steel cross with four arms, each instrumented with a foil-gauge full bridge; the cross arms deflect torsionally under joint torque while remaining stiff against shear and bending loads (which would corrupt the torque reading).
Specifications: 176 N·m full scale for the elbow-and-distal joints (J1–J4), 320 N·m for J0; ~16 000 N·m/rad stiffness; 0.2 % FS precision after factory calibration; per-joint update rate 1 kHz over the joint’s internal serial bus (Sercos-III). The joint controller implements Cartesian impedance control without a wrist F/T sensor: end-effector wrench is reconstructed from the joint-torque vector via the manipulator Jacobian, w_ext = J^{-T} · (τ_meas − τ_gravity − τ_friction). This is the canonical example of “torque-controlled robot” architecture and is the basis for KUKA’s “Smart Robotics” cobot offerings.
Meta AI DIGIT tactile fingertip (2020)
Lambeta, Chou, Tian, et al. (IEEE RAL 2020) introduced DIGIT, a $15-BOM fingertip-scale optical tactile sensor derived from MIT’s GelSight (Yuan, Dong, Adelson 2017). Three small LEDs (red / green / blue) illuminate the inside of a soft elastomer dome with a thin reflective skin; a 320×240 pixel global-shutter camera (OmniVision OV7251) images the deformation at 60 fps. The colour-coded illumination enables photometric stereo: the gradient of intensity at each pixel and each wavelength encodes the local surface normal of the deformed gel.
DIGIT’s contribution is engineering: every component is sub-$5, the firmware and CAD files are open-source at digit.ml, and the sensor mounts on a standard Allegro / Franka / Yale-OpenHand fingertip footprint. The associated PyTorch library tacto provides a simulator. Used in dozens of research papers since 2020 for in-hand manipulation (See, Kalashnikov 2021), insertion (Hogan 2020), and texture classification.
da Vinci Surgical Xi haptic feedback (Intuitive Surgical, 2014–)
The da Vinci Surgical Xi end-effector (“EndoWrist”) is a 7-DOF cable-driven instrument with a single-axis strain-gauged grip force transducer at the tool jaw and a tip-side 3-axis force sensor in some research variants (the production system is well known to limit haptic feedback through the cables, and Intuitive has published on the trade). Force feedback is rendered to the surgeon’s master console via voice-coil actuators at each fingertip control.
The case study illustrates a recurring theme in surgical robotics: friction in the cable drives between actuator and tool dominates the force-signal-to-noise ratio, so even a high-precision tip sensor delivers limited rendered haptics. Newer designs (Vicarious Surgical, Moon Surgical Maestro 2024) use direct-drive joints near the tool to recover force-feedback bandwidth.
DLR Hand II tactile fingertips (DLR/Schunk, 2010s)
Each of DLR Hand II’s fingertips integrates a 6-axis fingertip force sensor plus a distributed tactile array of 12 piezoresistive cells under the silicone skin. Total per-finger sensor count: 1 × 6-axis F/T + 12 tactile + 3 motor encoders + 3 joint encoders. The hand demonstrated dexterous in-hand manipulation tasks in the 2010s and informed the design of Awiwi and the David humanoid hand.
NASA Robonaut 2 hand tactile sensing (ISS, 2011–2018)
NASA’s Robonaut 2 (Diftler et al, ICRA 2011) was the first humanoid to operate inside the International Space Station. Each of its five-fingered hands (12 DOF per hand) integrated six-axis fingertip F/T sensors at the tip of each finger (custom strain-gauged transducers, ~0.05 N resolution) plus tactile arrays along each phalange. Total per-hand sensor count: 5 × 6-axis F/T + ~40 tactile cells. The tactile arrays used capacitive cells in a Kevlar-reinforced silicone glove to survive astronaut handling.
The case study illustrates the integration challenge as much as the sensor technology: routing 80 sensor channels through a 7-DOF arm, dealing with cable strain across joints, EMI from neighbouring motors, and thermal cycling between cold stowage (~5 °C) and operational (~30 °C). Robonaut’s experience drove NASA Valkyrie (R5) toward fewer-but-better-integrated sensors: 6-axis wrist F/T per arm + simplified strain-gauge fingertips.
Cobot peg-in-hole demo using only joint-torque sensors (Franka Panda, 2017)
Haddadin, Albu-Schäffer et al (2017, Springer Tracts on Advanced Robotics) demonstrated peg-in-hole insertion on a 1.5 mm clearance hole with no wrist F/T sensor — purely from joint-torque-derived Cartesian-wrench estimates. The trick: high-bandwidth (≥ 1 kHz) joint-torque sensing plus an accurate dynamic model (so the gravity, friction, and inertial torques can be subtracted from the measured torque). The accuracy of the estimated wrench was within 0.3 N RMS on the lateral axes, sufficient to detect contact at the chamfer and switch into search mode.
Shadow Robot Dexterous Hand (Shadow Robot Company, 2005–)
The Shadow Hand is a 24-DOF anthropomorphic hand widely used in dexterous-manipulation research. Force/tactile instrumentation has evolved across generations:
- Mk.III (2005–2010s): pneumatic muscles with embedded pressure sensors at each muscle; finger-tip force from a 6-axis strain-gauged sensor; tactile arrays optional.
- C6M (2015–): electric-motor-driven version; per-tendon force sensing via in-line load cells (1 mN resolution); BioTac tactile fingertips standard, optionally replaced with GelSight or DIGIT.
- Modular Grasper (2020–): simplified 3-finger configuration aimed at sim-to-real research; tactile data published as a ROS 2 topic at 1 kHz.
The Shadow Hand is the de-facto research benchmark for tactile manipulation: it has been used in OpenAI’s Rubik’s cube manipulation (Akkaya 2019), DeepMind’s dexterous-manipulation papers, and most recent ReSkin / DIGIT publications. The sensor cost per hand exceeds €100 000; the value to research is the integration density and software stack.
Polishing application with hybrid position/force control (industrial cell)
A representative industrial deburring cell uses a Fanuc M-20iA arm fitted with an ATI Gamma wrist F/T sensor (range ±130 N / ±10 N·m, resolution 0.025 N / 0.001 N·m) holding a pneumatic spindle with a bristle-disc brush. The control architecture decomposes Cartesian directions into position-controlled (along the workpiece surface, following a CAD path) and force-controlled (normal to the surface, holding 20 N contact).
Wrench measurements are taken at 1 kHz with hardware low-pass at 200 Hz to suppress the spindle’s 8 000 rpm vibration spectrum. The normal-force PID gains are: K_p = 5 mm/(N·s), K_d = 0.5 mm/N, K_i = 1 mm/(N·s²) — tuned so the response is well-damped and the brush rides the surface contour at travel speed 20 mm/s. End-effector compliance K = F_normal / Δz is set to ~1000 N/m, soft enough to compensate the ±0.5 mm CAD-to-real geometric error without overshoot.
The case illustrates the rule: the smallest force you can stably command is ~3–5× the sensor noise floor, and the loop bandwidth is set by the structural-mode of the workpiece + tool fixture (~30 Hz here, with the loop running at 1/5 of that, ~6 Hz on the normal axis).
Prosthetic hand sensorisation (Open Bionics Hero Arm, 2018–)
The Hero Arm is a commercial myoelectric prosthetic hand for upper-limb amputees, manufactured by Open Bionics (Bristol, UK). Force sensing is intentionally minimal — the arm uses an EMG-driven open-loop grip-force command from the user, with a single fingertip pressure sensor (FSR-style) per finger as a contact-confirmation switch. The simplicity is deliberate: every additional sensor adds cost, weight, and a failure mode in a consumer-priced ($10–25k) device.
Research prosthetic hands (e.g. DARPA Modular Prosthetic Limb, JHU APL 2014; Vincent Evolution4, 2020s) extend tactile feedback dramatically: 6-axis fingertip F/T, distributed pressure arrays, temperature, vibration — all fed back to the user via vibrotactile, electrotactile, or peripheral-nerve stimulation. The case illustrates the long path from research-grade sensorisation to consumer-grade integration: a sensor that works on a benchtop must survive sweat, impact, dropped objects, washing, and 5–10 years of daily wear before it ships in a consumer arm.
Boston Dynamics Atlas hand (2024 generation)
The third-generation Atlas humanoid (electric, announced April 2024) uses a custom 3-finger gripper with integrated 6-axis fingertip F/T sensors at each finger tip. Boston Dynamics has not published full specifications, but inferred from demo videos: sub-newton force resolution, ~100 Hz update rate, and likely a strain-gauge or piezoresistive transducer behind the rubber fingertip pad. The whole-arm reference uses a combination of the fingertip F/T plus joint-torque sensors at each of the 7 arm joints (similar architecture to the iiwa or Tesla Optimus) for compliant manipulation and safe interaction with the environment.
9. Cross-references
[[Robotics/sensors-pose-motion]]— companion proprioception note covering encoders, resolvers, and IMUs.[[Robotics/sensors-perception]]— exteroceptive sensors (LiDAR, cameras, ToF) that complement contact sensing.[[Robotics/impedance-control]]— closed-loop control architectures that consume the force/torque measurements from this note.[[Robotics/end-effectors]]— gripper integration: which tactile sensor fits which fingertip geometry.[[Robotics/manipulator-design]]— wrist sensor placement, joint sensor packaging, structural-mode interactions.[[Robotics/dynamics-rigid-body]]— Jacobian-based mapping from joint torques to end-effector wrench.[[Robotics/bayesian-estimation]]— filtering and fusion of force, tactile, and proprioceptive signals.[[Engineering/op-amps]]— bridge instrumentation amplifiers and signal conditioning.[[Engineering/semiconductor-devices]]— Hall-effect, piezoresistive, and MEMS device physics behind tactile sensors.[[Engineering/electromagnetics-engineering]]— capacitive sensing principles for tactile arrays.[[Languages/Tier3/robotics-control]]— ROS 2 / DDS topics carrying WrenchStamped and TactileMap messages.
Cost-vs-capability frontier (snapshot 2026)
| Capability | Floor cost | Reasonable cost | Premium cost |
|---|---|---|---|
| 6-axis wrist F/T, 0.5 N / 0.005 N·m | $4 000 (Robotous RFT80) | $7 000 (ATI Mini40) | $12 000 (Bota SensONE EtherCAT) |
| 6-axis wrist F/T, 0.05 N / 0.0005 N·m | n/a | $5 000 (ATI Nano17) | $9 000 (ATI Nano25) |
| Joint-torque sensor for cobot integration | $1 500 (basic strain-gauge) | $4 000 (Robotous RTS) | $10 000 (custom Sensitec) |
| Tactile fingertip array, 16+ cells | $50 (DIY DIGIT) | $500 (XELA uSkin) | $5 000 (BioTac, when available) |
| Whole-arm safety skin, 1 m² | n/a | $5 000 (custom capacitive) | $30 000 (KUKA Cybertouch) |
These are list-price ranges; volume discounts of 20–40 % are typical for OEM integration of 100+ units per year.
Procurement timeline reality
A practical note for project planning: lead times on premium F/T and joint-torque sensors are non-trivial. Typical 2024–2026 lead times:
- ATI Mini40 / Mini45: 6–10 weeks from order to ship; longer for custom calibration ranges.
- ATI Gamma / Delta: 8–14 weeks.
- Bota SensONE / Rokubi: 4–8 weeks (smaller European volume, faster cycle).
- SCHUNK FT: 8–12 weeks.
- OnRobot HEX: 2–6 weeks (cobot accessory volume, kept in stock).
- Custom joint-torque sensors: 12–20 weeks plus calibration.
- GelSight Mini / DIGIT: 2–4 weeks from order; DIGIT is open-source and can be self-fabricated for ~$50 BOM in 1 week if local machining is available.
Plan accordingly. A project that needs sensor-in-hand before the next milestone should order at least 3 months ahead, with a confirmed BOM and electrical-interface spec.
Looking ahead
Three trends shape the 2025–2030 force/tactile landscape:
- Whole-body skin coverage for humanoids. Tesla Optimus, Figure 02, 1X Neo, Apptronik Apollo, Boston Dynamics Atlas all integrate distributed skin-class sensors over arms and torso. Cost trajectory drives adoption: AKM, Melexis, and ams Osram are bringing 3-axis magnetic Hall ICs below 5–15 per tactile cell.
- Learning-based tactile inference. End-to-end CNN/transformer architectures decode raw tactile-array images directly into object identity, pose, grasp quality, and slip onset without explicit force-pressure calibration. ReSkin and DIGIT both released large open-source datasets (10⁵+ contact events) for training. The trade-off: black-box predictions that are difficult to integrate into safety-rated systems but excellent for manipulation policies.
- High-bandwidth event-driven tactile cameras. Inspired by silicon-retina event cameras (Prophesee, iniVation), tactile-event sensors emit asynchronous pixel-level pressure changes at sub-ms latency rather than fixed frame rates. Research-stage in 2024–25; early commercial offerings expected by 2027.
10. Citations
- Murray, R. M., Li, Z., & Sastry, S. S. (1994). A Mathematical Introduction to Robotic Manipulation, CRC Press, ch. 5 (force/torque, Jacobian transposes, screw theory).
- Siegwart, R., Nourbakhsh, I. R., & Scaramuzza, D. (2011). Introduction to Autonomous Mobile Robots (2nd ed.), MIT Press, ch. 4 (contact sensors).
- Lumelsky, V. J., Shur, M. S., & Wagner, S. (2001). “Sensitive skin.” Communications of the ACM, 44(8), pp. 24–26, and IEEE Sensors Journal, 1(1), pp. 41–51. Foundational vision for whole-body robot skin.
- Yousef, H., Boukallel, M., & Althoefer, K. (2011). “Tactile sensing for dexterous in-hand manipulation in robotics — A review.” Sensors and Actuators A: Physical, 167, pp. 171–187. DOI 10.1016/j.sna.2011.02.038.
- Albu-Schäffer, A., Haddadin, S., Ott, C., Stemmer, A., Wimböck, T., & Hirzinger, G. (2007). “The DLR lightweight robot — design and control concepts for robots in human environments.” Industrial Robot: An International Journal, 34(5), pp. 376–385. DOI 10.1108/01439910710774386. Original architecture for joint-torque-based impedance control on the LWR-III / KUKA iiwa lineage.
- Lambeta, M., Chou, P.-W., Tian, S., Yang, B., Maloon, B., Most, V. R., Stroud, D., Santos, R., Byagowi, A., Kammerer, G., Jayaraman, D., & Calandra, R. (2020). “DIGIT: A novel design for a low-cost compact high-resolution tactile sensor with application to in-hand manipulation.” IEEE Robotics and Automation Letters, 5(3), pp. 3838–3845. DOI 10.1109/LRA.2020.2977257.
- Wettels, N., Santos, V. J., Johansson, R. S., & Loeb, G. E. (2008). “Biomimetic tactile sensor array.” Advanced Robotics, 22(8), pp. 829–849. DOI 10.1163/156855308X314533. BioTac multi-modal sensor.
- Fishel, J. A., & Loeb, G. E. (2012). “Sensing tactile microvibrations with the BioTac — Comparison with human sensitivity.” Proceedings of the IEEE RAS/EMBS BioRob, pp. 1122–1127. DOI 10.1109/BioRob.2012.6290741. Slip detection via spectral analysis.
- Yuan, W., Dong, S., & Adelson, E. H. (2017). “GelSight: High-resolution robot tactile sensors for estimating geometry and force.” Sensors, 17(12), 2762. DOI 10.3390/s17122762.
- Tomo, T. P., Schmitz, A., Wong, W. K., Kristanto, H., Somlor, S., Hwang, J., Jamone, L., & Sugano, S. (2018). “Covering a robot fingertip with uSkin: A soft electronic skin with distributed 3-axis force sensitive elements for robot hands.” IEEE Robotics and Automation Letters, 3(1), pp. 124–131. DOI 10.1109/LRA.2017.2734965.
- Mittendorfer, P., & Cheng, G. (2011). “Humanoid multimodal tactile-sensing modules.” IEEE Transactions on Robotics, 27(3), pp. 401–410. DOI 10.1109/TRO.2011.2106330. HEX-O-SKIN / CellulARSkin.
- Bhirangi, R., Hellebrekers, T., Majidi, C., & Gupta, A. (2021). “ReSkin: Versatile, replaceable, lasting tactile skins.” Conference on Robot Learning (CoRL) 2021. arXiv:2111.00071.
- Haddadin, S., Albu-Schäffer, A., & Hirzinger, G. (2017). “Robot collisions: A survey on detection, isolation, and identification.” IEEE Transactions on Robotics, 33(6), pp. 1292–1312. DOI 10.1109/TRO.2017.2723903.
- ATI Industrial Automation (2024). Mini40 / Mini45 / Nano17 6-axis force/torque sensor datasheets, system manual rev 9.
- SCHUNK GmbH (2024). FT / FTN / FTM Axia series product catalog rev 4.
- Bota Systems AG (2024). Rokubi / SensONE / MiniONE 6-axis F/T sensor product datasheets.
- Robotous (2023). RFT80-6A01 6-axis force/torque sensor datasheet rev 2.
- Tekscan Inc (2024). FlexiForce A201 / A301 / A401 standard force sensor datasheets, rev I.
- Pressure Profile Systems (2023). RoboTouch and FingerTPS tactile sensor product brochures.
- GelSight Inc (2024). GelSight Mini product page and datasheet.
- Texas Instruments (2023). ADS1262 / ADS1263 32-bit Σ-Δ ADC datasheet rev F.
- Analog Devices (2024). AD7124-8 Low-power Σ-Δ ADC datasheet rev D.
- ISO 8373:2021. Robots and robotic devices — Vocabulary.
- ISO 10218-1:2025. Robotics — Safety requirements — Part 1: Industrial robots.
- ISO/TS 15066:2025. Robots and robotic devices — Collaborative robots. Defines power-and-force-limiting (PFL) thresholds that drive joint-torque and contact-skin specifications.
- Diftler, M. A. et al. (2011). “Robonaut 2 — the first humanoid robot in space.” ICRA 2011, pp. 2178–2183. DOI 10.1109/ICRA.2011.5979830.
- Cutkosky, M. R., Howe, R. D., & Provancher, W. R. (2008). “Force and tactile sensors.” In Springer Handbook of Robotics (1st ed., ch. 19). DOI 10.1007/978-3-540-30301-5_20. Canonical handbook chapter.
- Howe, R. D., & Cutkosky, M. R. (1993). “Dynamic tactile sensing — Perception of fine surface features with stress rate sensing.” IEEE Transactions on Robotics and Automation, 9(2), pp. 140–151. DOI 10.1109/70.238278.
- Westling, G., & Johansson, R. S. (1984). “Factors influencing the force control during precision grip.” Experimental Brain Research, 53(2), pp. 277–284. The biological reference for human-fingertip slip detection thresholds.
- Stassi, S. et al. (2014). “Flexible tactile sensing based on piezoresistive composites — A review.” Sensors, 14(3), pp. 5296–5332. DOI 10.3390/s140305296.
- Dahiya, R. S., Metta, G., Valle, M., & Sandini, G. (2010). “Tactile sensing — From humans to humanoids.” IEEE Transactions on Robotics, 26(1), pp. 1–20. DOI 10.1109/TRO.2009.2033627.
- ISO 9409-1:2004. Manipulating industrial robots — Mechanical interfaces — Part 1: Plates. Defines the wrist mounting flange specification used by every commercial F/T sensor.
- ISO 13849-1:2023. Safety of machinery — Safety-related parts of control systems. PLd / Cat 3 architectural requirements applicable to safety-rated F/T and skin sensors.
- IEC 60068-2-6:2007. Environmental testing — Vibration (sinusoidal). Standard vibration profile for qualifying robotic sensors.
- IEC 60529:2013. Degrees of protection provided by enclosures (IP code). Used for IP67 / IP69K wet-environment F/T sensor rating.