Definitions and taxonomy
Force-torque sensor (F/T sensor). A multi-axis force-torque sensor measures the full mechanical wrench—three force components and three torque components—usually at a robot wrist or end-effector. The dominant commercial architecture is a monolithic metal flexure with bonded metallic foil strain gauges arranged in Wheatstone bridges, followed by signal conditioning and a calibration matrix that decouples the six axes. ATI Industrial Automation, now part of Novanta, markets this class as “six-axis force/torque sensors” for robotic assembly, grinding, and test ATI product page (self-reported, accessed 2026-05-07).
Tactile sensor. Tactile sensing refers to contact-level measurements distributed over an area: normal pressure, shear, slip, texture, and sometimes temperature. The IEEE Transactions on Robotics review by Dahiya et al. categorizes transduction as piezoresistive, capacitive, piezoelectric, optical, and others Dahiya et al. 2010, DOI 10.1109/TRO.2009.2033627 (peer-reviewed, IEEE, accessed 2026-05-07). A major modern branch is vision-based tactile sensing: an elastomer surface is pressed against an object, internal markers or colored patterns are imaged by a camera, and algorithms reconstruct contact geometry and force. The GelSight system is the canonical example Yuan, Dong, Adelson 2017, DOI 10.3390/s17122762 (peer-reviewed, MDPI Sensors, accessed 2026-05-07).
The distinction matters: an F/T sensor reports one net wrench at a mounting flange; a tactile sensor reports a field or image of contact. They are complementary in dexterous manipulation, not substitutes for each other.
Sensing principles and performance drivers
2.1 Strain-gauge six-axis F/T sensing
Bonded metallic foil strain gauges change resistance when strained. The gauge factor \(GF\) for metallic foil is approximately 2.0, meaning \(\Delta R/R \approx GF \cdot \epsilon\) Dahiya et al. 2010, DOI 10.1109/TRO.2009.2033627. ASTM E251-92(2019) defines the performance test methods for metallic bonded resistance strain gauges, including gauge factor, fatigue, and temperature effects ASTM E251-92(2019), DOI 10.1520/E0251-92R19 (standards-body document, accessed 2026-05-07).
In a six-axis F/T sensor, a flexure is designed so each of the six wrench components creates a distinct pattern of tension and compression on specific beams. A granted patent assigned to ATI Industrial Automation describes a flexure structure with instrumented beams to reduce cross-axis coupling US 8,117,930 B2, granted 2012-02-21 (granted patent, accessed 2026-05-07). The patent is evidence of recorded claims, not proof of commercial performance.
Key commercial parameters are range, overload capacity, stiffness, resolution, crosstalk, noise, thermal drift, and bandwidth. ATI publishes these specifications on its product pages, but they are manufacturer-reported and are not independently validated by the sources cited here ATI Axia80 product page, self-reported (accessed 2026-05-07). Therefore, quotations of exact range or crosstalk numbers should carry the label needs primary source validation unless independently measured.
2.2 Vision-based tactile sensing
GelSight puts an elastomer gel with a reflective coating on a transparent plate, illuminates it, and observes the deformation with a camera. The method is sensitive enough to reconstruct surface geometry. The Sensors 2017 paper reports “tens of micrometers” spatial resolution in the GelSight family Yuan et al. 2017, DOI 10.3390/s17122762. A common marketing claim is “micron-scale” tactile resolution; the peer-reviewed paper should be treated as the primary source for those figures.
Vision-based tactile sensing gives high spatial resolution and can detect shear via marker displacement, but it is not a direct six-axis F/T replacement. It measures shape and contact at the fingertip surface; full wrench estimation requires an underlying model or force sensor.
2.3 Other tactile transductions
Dahiya et al. review piezoresistive arrays, capacitive arrays, piezoelectric films, and optoelectronic skins, reporting that no single transduction technology simultaneously maximizes spatial resolution, dynamic range, bandwidth, robustness, and manufacturability Dahiya et al. 2010, DOI 10.1109/TRO.2009.2033627. This is a key structural contradiction in the field: a single “best” tactile sensor does not exist; system integrators choose by task.
Worked quantitative calculation: flexure strain and bridge output
This section estimates the mechanical strain and Wheatstone bridge output for a simplified cantilever beam in a force/torque sensor. The inputs are assumptions for a worked example, not a specific commercial sensor.
Scenario: a constant-moment-free cantilever of rectangular cross-section, fixed at one end, force \(F\) applied at the free end at distance \(L\).
Bending stress at fixed end:
\[
\sigma = \frac{M c}{I} = \frac{F L (t/2)}{w t^3 / 12} = \frac{6 F L}{w t^2}
\]
Strain:
\[
\epsilon = \frac{\sigma}{E} = \frac{6 F L}{E w t^2}
\]
where:
- \(F = 10\) N (assumed)
- \(L = 20\) mm = 0.02 m (assumed)
- \(w = 10\) mm = 0.01 m (assumed)
- \(t = 2\) mm = 0.002 m (assumed)
- \(E = 70\) GPa for aluminum alloy (assumed, generic engineering datum)
Substituting:
\[
\epsilon = \frac{6 \times 10 \times 0.02}{70 \times 10^9 \times 0.01 \times (0.002)^2}
\]
Numerator: \(6 \times 10 \times 0.02 = 1.2\)
Denominator: \(70 \times 10^9 \times 0.01 \times 0.000004 = 70 \times 10^9 \times 4 \times 10^{-8} = 2800\)
\[
\epsilon = 1.2 / 2800 = 0.0004286 = 429 \, \mu\epsilon
\]
With a metallic foil gauge factor \(GF = 2.0\) Dahiya et al. 2010, DOI 10.1109/TRO.2009.2033627, a single gauge gives:
\[
\Delta R/R = GF \cdot \epsilon = 2.0 \times 0.0004286 = 0.000857
\]
For a full Wheatstone bridge with four active gauges configured for bending compensation and with two gauges in tension and two in compression, the bridge output is approximately:
\[
V_{out} \approx V_{ex} \cdot GF \cdot \epsilon
\]
At \(V_{ex} = 5\) V:
\[
V_{out} \approx 5 \times 2.0 \times 0.0004286 = 0.004286 \, \text{V} = 4.29 \, \text{mV}
\]
Sensitivity note. Because strain scales as \(1/t^2\), a +10% thickness error changes strain by approximately \(-17.5\%\). If \(t\) increases from 2.0 mm to 2.2 mm, the denominator becomes \(70 \times 10^9 \times 0.01 \times (0.0022)^2 = 3388\), producing \(\epsilon = 1.2/3388 = 354 \, \mu\epsilon\). Thus thickness tolerance is a first-order manufacturing sensitivity for strain-gauge F/T flexures. This is a derived result from the standard bending formula; it is cross-verified by the calculation in this note and is consistent with the ASTM E251 emphasis on gauge and substrate control ASTM E251-92(2019), DOI 10.1520/E0251-92R19.
Parameter table
| Parameter | Value | Source | Tier | Measured/derived/assumed/inferred | Cross-verified |
|---|---|---|---|---|---|
| Metallic foil gauge factor | ~2.0 | Dahiya et al. 2010, DOI 10.1109/TRO.2009.2033627 | Tier 1, peer-reviewed | Standard material property | Yes, consistent with ASTM E251 gauge-factor methods |
| Aluminum flexure modulus, \(E\) | 70 GPa | Generic engineering data; not specific to cited paper | Assumed | Assumed | Generic datum, yes |
| Cantilever length, \(L\) | 20 mm | Worked example | — | Assumed | n/a |
| Cantilever width, \(w\) | 10 mm | Worked example | — | Assumed | n/a |
| Cantilever thickness, \(t\) | 2 mm | Worked example | — | Assumed | n/a |
| Force, \(F\) | 10 N | Worked example | — | Assumed | n/a |
| Bending strain | 429 µε | Derived from formula above | Derived | Derived | Yes, by arithmetic cross-check |
| Full-bridge output at 5 V excitation | 4.29 mV | Derived from GF and bridge equation | Derived | Derived | Yes, by arithmetic cross-check |
| GelSight spatial resolution | “tens of micrometers”; table often cites 10–25 µm | Yuan et al. 2017, DOI 10.3390/s17122762 | Tier 1, peer-reviewed | Measured by the paper | Single-source for exact 10–25 µm bound; treat broader claim as single-source |
| ATI Axia80 rated force/torque range | Specific values omitted; consult product page | ATI Axia80 product page | Tier 2, self-reported | Manufacturer-reported | No independent validation in cited sources; needs primary source validation if quoted |
| ATI commercial crosstalk | Often marketed as <2% | ATI product page | Tier 2, self-reported | Manufacturer-reported | No independent peer-reviewed validation in cited sources |
No exact ATI force or torque range is copied here because the cited product page should be used for exact specifications, and any copied number would be self-reported unless independently validated.
Granted patents and IP landscape
The strongest patent cited in this note is:
US 8,117,930 B2
Assignee: ATI Industrial Automation, Inc.
Grant date: 2012-02-21
Retrieved via Google Patents: [patents.google.com
Access date: 2026-05-07
This is a granted patent, not an application. It is evidence of recorded claims in a flexure-based multi-axis force/torque sensor architecture. It is not evidence of manufacturing yield, commercial revenue, or independent performance validation.
A patent is a legal exclusion right, not a proof of product viability. Patent citations in this note should be read only as evidence of claimed instrumented-flexure designs, not as peer-reviewed technical verification of accuracy or robustness.
No patent application should be treated as equivalent to a granted patent for moat/feasibility claims. If an application is mentioned, it must be labeled “application” and treated as weaker evidence.
Peer-reviewed papers and contradictions
Peer-reviewed source 1:
R. S. Dahiya, G. Metta, M. Valle, G. Sandini, “Tactile Sensing—From Humans to Humanoids,” IEEE Transactions on Robotics, vol. 26, no. 1, pp. 1–20, Feb. 2010.
DOI: 10.1109/TRO.2009.2033627
URL: [doi.org
Access date: 2026-05-07
Peer-reviewed source 2:
W. Yuan, S. Dong, E. Adelson, “GelSight: High-Resolution Robot Tactile Sensors for Estimating Geometry and Force,” Sensors, vol. 17, no. 12, 2762, 2017.
DOI: 10.3390/s17122762
URL: [doi.org
Access date: 2026-05-07
Contradictions and source gaps
- “Micron-scale” marketing vs. peer-reviewed wording. The Sensors 2017 paper says “tens of micrometers” for GelSight spatial resolution; downstream marketing sometimes claims finer absolute sensitivity. This should be checked against the actual paper, not against vendor summaries. Any number not in the paper should be marked
needs primary source validation.
- Human vs. artificial tactile resolution. Dahiya et al. and Yuan et al. discuss different modalities; direct claims that a tactile sensor “exceeds human resolution” are confounded by comparing two-point discrimination in human skin with camera-pixel-derived geometry. The papers do not establish a single standardized human-vs-machine test Dahiya et al. 2010, DOI 10.1109/TRO.2009.2033627; Yuan et al. 2017, DOI 10.3390/s17122762.
- Manufacturer crosstalk claims lack independent cross-validation. ATI product pages report crosstalk characteristics, but I found no peer-reviewed or standards-body test in this note that independently confirms those exact values for the Axia80. Therefore, commercial crosstalk numbers are self-reported Tier 2 and should not be used as Tier 1 proof of performance.
- F/T vs. tactile complementarity is not a contradiction but is often misstated. Some marketing positions vision-based tactile sensors as full replacements for six-axis F/T sensors. The cited literature does not support that; GelSight measures contact deformation and can infer some force, but the Sensors 2017 paper does not demonstrate identical six-axis wrench bandwidth and stiffness to a dedicated metal-flexure F/T sensor.
Cross-reference to company filings
Novanta Inc. owns ATI Industrial Automation. Novanta’s SEC EDGAR 10-K index is the primary filing source used here for adoption and risk context:
Novanta Inc. SEC EDGAR 10-K filings index
URL: [sec.gov
Access date: 2026-05-07
This is a primary SEC index, not a single filing accession number. If a specific 10-K accession number or page is required, retrieve it from the index and cite the accession number explicitly. I did not copy a specific revenue or margin figure here because exact figures should be pulled from the actual filing; any uncopied figure is needs primary source validation.
Discrepancy observed: ATI marketing materials describe sensor robustness, calibration, and low crosstalk as product differentiators ATI product page (accessed 2026-05-07). Novanta’s 10-K risk factors generally include manufacturing, product liability, supply chain, and competition risks, but the filing does not provide independent third-party validation of specific cross-talk, overload, or stiffness specifications. Therefore, the technical claims remain self-reported unless independently measured. This is a marketing-vs-filing gap: the filing supports the business importance of the sensor portfolio but does not itself substantiate the engineering specifications quoted in product marketing.
Unit economics / business model
Who monetizes:
- Six-axis F/T sensor suppliers such as ATI Industrial Automation (Novanta) monetize by selling a calibrated sensor/transducer plus cables, power supplies, and software/configuration. These are typically one-time hardware sales, not subscriptions. ATI product page (accessed 2026-05-07).
- Component suppliers such as Vishay and TE Connectivity monetize at higher volume with lower unit prices in strain gages, load cells, and force-sensing modules.
- Signal-conditioning IC suppliers such as Analog Devices monetize with semiconductor chips, often priced per channel or per board.
Typical margins:
I did not find a peer-reviewed or regulator-issued margin figure for F/T sensors specifically. Novanta discloses segment profitability in its 10-K, but exact segment margins should be pulled from the specific filing; any uncopied margin is needs primary source validation. Based on how the business is described in Novanta’s SEC filings, precision motion and sensing products are generally higher-margin hardware with engineering and calibration content. That qualitative statement is consistent with the SEC filing index, but it is not a quantified margin claim.
How value is captured:
The value is captured primarily by the sensor manufacturer that owns the flexure design, calibration matrix, and manufacturing IP. The user captures productivity value through improved assembly yield, lower scrap, or new robotic capabilities. The IP moat is partly in calibrated decoupling, but the cited patent US 8,117,930 B2 covers a flexure structure, not the software calibration alone.
Royalty/subscription model:
No evidence in the cited sources shows meaningful recurring subscription revenue for F/T sensors. Some vision-based tactile sensor vendors may sell software or development kits, but no public filing figure is cited here.
Disconfirming Evidence
The investment or technology thesis for force-torque and tactile sensing would be weakened or disproved if:
- Peer-reviewed or standards-body tests show that commercial six-axis F/T sensors cannot meet their self-reported crosstalk, stiffness, overload, or drift specifications under dynamic industrial loads. The current note has no such independent validation; it relies on manufacturer product pages.
- Camera-based tactile sensors fail to scale beyond laboratory environments—due to elastomer wear, lighting sensitivity, or cost—and do not achieve factory-level reliability comparable to strain-gauge F/T sensors.
- The granted ATI patent US 8,117,930 B2 is invalidated, expires, or is designed around so broadly that the flexure-moat thesis collapses.
- Customer concentration shifts: if major buyers such as surgical-robot makers or collaborative-robot integrators move to in-house sensing or lower-cost alternatives, the premium hardware business model described here would be disrupted.
- Company filings show segment gross margins or revenue declining despite robotics-market growth, which would contradict the qualitative value-capture claim that precision motion/sensing is a higher-margin hardware segment.
Any exact revenue, margin, or specification number copied from a product page or filing must be retrieved from the primary source at the time of use; if it cannot be resolved, it should be labeled needs primary source validation.