Precision motion control / force-torque sensing for humanoid and cobot

Executive summary

Force–torque (F/T) sensing is a load-bearing input for precision motion control in collaborative robots (cobots) and humanoid robots. A six-axis F/T transducer measures three orthogonal force components and three orthogonal torque components in a single calibrated reference frame. In humanoids, these sensors are placed at the ankle/foot, wrist, hip, or end-effector to estimate ground-contact wrenches, control contact tasks, and close an impedance/admittance control loop. In cobots, F/T sensing enables hand-guiding, assembly, polishing, screw-driving, force-limited contact, and human–robot collaboration.

The foundational sensing principle is well documented in the granted patent US 4,094,192, which describes a six-degree-of-freedom force and torque measurement apparatus using strain-sensitive elements arranged to resolve the six independent wrench components. The optimal mechanical design of such transducers was formalized in the peer-reviewed literature by Bicchi (1992), who articulated the design objective of minimizing cross-talk and maximizing force-sensing isotropy. These are enduring design goals: a six-axis F/T transducer should have low condition number in its calibration matrix, high overload tolerance, high stiffness, and repeatable output under temperature drift and mechanical load.

The commercial value is concentrated in precision transducer suppliers, robot OEMs, signal-chain semiconductor vendors, and advanced-robot developers. The public-company exposure spans Novanta (ATI Industrial Automation), Teradyne (Universal Robots), Tesla, Intuitive Surgical, Analog Devices, Texas Instruments, Vishay Precision Group, Honeywell, Rockwell Automation, Fanuc ADR, and Yaskawa ADR. Each connection is listed with primary-source URLs later in this note.

Physics and sensing architectures

2.1 What a six-axis F/T sensor measures

A six-axis F/T sensor outputs a wrench:

\[

w = [F_x, F_y, F_z, T_x, T_y, T_z]^T

\]

where \(F_x, F_y, F_z\) are forces in Newtons and \(T_x, T_y, T_z\) are torques in Newton-meters or Newton-millimeters. The sensor body is usually an elastic element, often a spoke, Maltese-cross, cylinder, or Stewart-platform-like structure. Strain gauges are bonded to regions that flex in predictable ways under each of the six wrench components. The earliest granted patent covering this architecture is US 4,094,192, titled “Method and apparatus for six degree of freedom force and torque measurement,” granted June 13, 1978 [S1]. It describes strain-responsive transducers arranged to separate the six components.

The later peer-reviewed design criterion by Bicchi explains why naive mountings of six gauges are not sufficient: the transformation from raw gauge signals to Cartesian wrench must be well-conditioned, or small noise in one channel can contaminate force and torque estimates [S2]. This is the “mechanical decoupling” requirement. Modern commercial sensors are marketed around the same requirement: ATI describes its six-axis sensors as resolving “all components of force and torque in a single transducer” [S3], and Bota Systems describes its sensors as “full 6-axis force torque sensing” with low noise and drift [S4].

2.2 Strain gauge versus alternative transduction

The dominant commercial architecture is metal-foil strain gauge on a machined flexure. The gauge factor of a constantan foil gauge is approximately 2.0, meaning fractional resistance change is about twice the mechanical strain [S5]. Semiconductor/silicon gauges have much higher gauge factor, often on the order of 50–150, but they are more temperature-sensitive and often require more careful compensation. Piezoelectric sensors are common in dynamic force measurement but have poor DC performance; capacitive and optical force sensors offer alternatives but have not displaced strain-gauge six-axis sensors in mainstream cobot/humanoid applications.

The sensing chain is typically:

1. Elastic flexure converts wrench to local strain.
2. Strain gauges convert strain to resistance change.
3. Wheatstone bridge converts resistance change to differential voltage.
4. Instrumentation amplifier/ADC converts voltage to digital counts.
5. Firmware applies a 6×6 calibration matrix to recover wrench components.
6. The host controller closes the force/motion loop.

This chain is visible across precision-instrumentation and semiconductor supplier documentation, including Analog Devices’ robotics signal-chain pages [S6] and Texas Instruments’ robotics applications pages [S7]. The signal-chain requirement is not trivial: microvolt-level bridge outputs must be amplified and digitized with enough dynamic range to capture both delicate contact forces and high overloads.

Why precision motion control depends on F/T sensing

3.1 Humanoid robots

A humanoid robot must estimate contact wrenches at the feet to control balance and impulsive ground reaction. Without trustworthy ankle/foot F/T signals, a bipedal controller cannot distinguish between a firm stance, a slipping contact, or an unexpected push. The same issue appears in the wrist/end-effector: grasping and tool use require contact force feedback to avoid crushing objects or losing grip. Tesla’s public Optimus demonstrations emphasize articulated hands, actuator integration, and sensor feedback [S8]. The company’s 10-K separately warns that the Optimus program may face development delays, manufacturing complexity, and commercial uncertainty [S9]. Those two source classes — a marketing/demo page and a filed risk section — are not necessarily contradictory, but they show the difference between demonstrated prototypes and filed risk disclosure. The 10-K does not certify that the sensors are production-qualified or that the cost curve will support volume manufacturing.

3.2 Collaborative robots

For cobots, force/torque sensing is used for:

• Hand-guiding, where a human moves the robot arm by pushing on it and the controller treats the F/T signal as a command input.
• Insertion and assembly tasks that require active compliance rather than position-only control.
• Polishing, grinding, deburring, screw-driving, and other contact processes in which force repeatability is the manufacturing variable of interest.
• Safety and force limiting in direct human–robot contact.

Universal Robots markets this class of capability on its e-Series and newer arms, which integrate joint torque sensing and support external F/T sensors [S10]. OnRobot and Robotiq sell external six-axis F/T sensors designed to bolt onto cobot wrists or tool flanges [S11][S12]. The F/T sensor is therefore not always a single “sensor product”; it can be a joint-level torque ring, an end-effector wrench sensor, or a force-limited actuator property.

Performance metrics, calibration, and signal chain

4.1 Key performance parameters

Relevant vendor specifications usually include:

• Nominal force and torque ranges per axis.
• Single-axis and multi-axis overload limits.
• Resolution in counts or N/N·mm.
• Crosstalk between axes.
• Hysteresis and nonlinearity.
• Output noise density or noise-limited resolution.
• Stiffness and resonant frequency.
• Temperature drift and compensation.
• Digital output protocol: EtherCAT, EtherNet/IP, PROFINET, CAN, USB, etc.

A key point from the Bicchi paper is that crosstalk is not only an electronics problem; it is primarily a mechanical design problem. A poorly conditioned flexure can amplify noise by an order of magnitude before any digital filtering is applied [S2].

4.2 Calibration matrix

The raw strain-gauge vector \(s = [s_1, s_2, \ldots, s_n]^T\), with \(n \ge 6\), is mapped to a wrench by a calibration matrix \(C\):

\[

w = C s

\]

For a six-channel design, \(C\) is 6×6. For redundant designs, \(C\) is 6×n and may be solved by least squares during calibration. The conditioning of \(C\) determines noise amplification. Bicchi formalized this as a design objective: the sensor structure should approach isotropic response and low condition number across the wrench space [S2].

Worked quantitative calculation: Wheatstone bridge sensitivity

This calculation uses the standard strain-gauge bridge equation, as presented in Vishay Micro-Measurements strain-gauge technical notes [S5]. It is a first-order linear approximation for a quarter-bridge.

5.1 Formula

For a single active gauge in a quarter bridge:

\[

V_{out} \approx V_{ex} \cdot \frac{GF \cdot \varepsilon}{4}

\]

Where:

• \(V_{ex}\): bridge excitation voltage, V
• \(GF\): gauge factor, dimensionless
• \(\varepsilon\): mechanical strain, dimensionless, typically expressed in microstrain (\(\mu\varepsilon\))
• \(V_{out}\): differential bridge output, V

5.2 Inputs

5.3 Calculation

\[

V_{out} = 5.0 \cdot \frac{2.0 \cdot 0.001}{4}

\]

\[

V_{out} = 5.0 \cdot 0.0005 = 0.0025\ V = 2.5\ mV

\]

5.4 Sensitivity and ADC resolution

If the flexure is designed so that 1,000 \(\mu\varepsilon\) corresponds to full-scale force \(F_{FS}\), then bridge sensitivity is:

\[

S = \frac{2.5\ mV}{F_{FS}}

\]

For a hypothetical design where \(F_{FS}=100\ N\):

\[

S = 25\ \mu V/N

\]

A 24-bit ADC with a ±5 V differential full-scale range has an ideal LSB of:

\[

LSB = \frac{10\ V}{2^{24}} = 0.596\ \mu V

\]

Ideal arithmetic resolution would be:

\[

\Delta F_{ideal} = \frac{0.596\ \mu V}{25\ \mu V/N} \approx 0.024\ N

\]

This is not the practical force resolution. Real noise from the strain gauge, junction thermal noise, amplifier noise, ADC noise, and mechanical vibration dominate. The calculation demonstrates that the electronic bit depth is not the limiting factor in modern F/T sensing; the mechanical and analog front end are the constraint.

5.5 Sensitivity check

If gauge factor changes by +1% (from 2.0 to 2.02), output changes from 2.500 mV to 2.525 mV, or +1%. This shows why shunt calibration and temperature compensation are necessary in precision F/T transducers. The cited patent US 4,094,192 provides an early description of strain-sensitive elements and bridge arrangements [S1]. Company documentation from ATI and Bota emphasizes calibration and noise performance because commercial users buy calibrated force/torque data, not raw millivolt signals [S3][S4].

5.6 Parameter table

Patent and IP landscape

6.1 A granted patent is not proof of performance

The strongest patent cited here is US 4,094,192, granted June 13, 1978 [S1]. It is a foundational six-degree-of-freedom force and torque measurement method. However, it is a granted patent, not an application, and patents are evidence of recorded claims, not proof of commercial performance. The patent long predates current humanoid and cobot market conditions, and its commercial relevance is historical/landscape-level rather than a live moat.

Modern F/T players hold later patents covering flexure geometry, overload protection, temperature compensation, and calibration methods. This note does not enumerate those patents unless independently resolved. Any unverified patent family counts or “core patent” assertions must be marked needs primary source validation if they cannot be traced to a grant number.

6.2 Patent moat versus manufacturing moat

A six-axis F/T sensor is a precision mechanical + calibration product. The moat is often stronger in manufacturing tolerances, calibration infrastructure, and robot-ecosystem integration than in a single patent. Bicchi’s paper shows that the mechanical optimization problem has been understood for decades [S2]; current commercial differentiation is therefore partly in sensor-specific ASIC/integration, digital interfaces, and software drivers.

Peer-reviewed and technical literature

The primary peer-reviewed source for optimal multi-axis force sensor design is Bicchi’s 1992 article in Robotics and Autonomous Systems [S2]. The paper addresses the central design problem: how to choose geometry and gauge placement so the transducer provides well-conditioned estimates of all six wrench components. It remains a structural reference because it ties mechanical design to estimation quality.

A limitation of this source for a 2025 technology note is age: it does not capture current digital calibration methods, EtherCAT/EtherNet/IP sensor interfaces, or humanoid-specific ankle/wrist packaging. For these, the note relies on Tier 2 company technical documentation from ATI, Bota Systems, OnRobot, Robotiq, and Universal Robots [S3][S4][S10][S11][S12]. These company sources are self-reported and should not be treated as independent performance verification.

Commercial adoption context and filing cross-check

8.1 ATI / Novanta

Novanta completed its acquisition of ATI Industrial Automation in 2020, bringing ATI’s force/torque sensors into a public-company structure [S13]. ATI’s products are marketed as robotic force/torque transducers for research, surgical robots, humanoids, and industrial automation [S3]. Novanta’s SEC filings disclose the ATI business as part of its precision-medical/industrial portfolio, but the filings do not break out ATI-specific sensor gross margin or unit volume in a way that permits margin verification from primary filing text alone. Any specific F/T sensor ASP or gross margin figure must be marked needs primary source validation.

8.2 Tesla Optimus

Tesla’s public robotics page describes Optimus as a general-purpose humanoid using Tesla-designed actuators and sensors [S8]. The company’s 10-K includes risk factors that warn against assuming the program will scale on the implied timeline or achieve commercial viability [S9]. The cross-reference finding is: demo disclosures are product/marketing claims; filed 10-K risk disclosures are the regulated record and do not certify production-qualified force/torque sensing or unit economics.

8.3 Universal Robots / Teradyne

Teradyne’s ownership of Universal Robots places it in the cobot value chain. Universal Robots’ e-Series and newer arms use joint torque sensing for force-limited motion, and its UR+ ecosystem supports F/T sensors from ATI, Robotiq, OnRobot, and others [S10]. Teradyne’s SEC filings report robotics segment performance but do not provide per-model BOM detail on F/T sensors. Therefore, the public company is exposed, but the specific sensor margin contribution is not visible from primary filings.

Unit economics / business model

9.1 Who monetizes

The monetization model for precision F/T sensing is primarily hardware sale, not subscription or royalty:

• Transducer makers sell the calibrated sensor body and, increasingly, integrated digital output electronics. Examples: ATI (Novanta), Bota Systems, OnRobot, Robotiq, Schunk, and others.
• Robot OEMs monetize F/T sensing as an integrated feature or optional accessory on cobot arms.
• Semiconductor vendors monetize the amplification, ADC, and interface silicon around the sensor.
• Software/tooling vendors monetize calibration, configuration, and force-control applications, though this is usually a smaller revenue stream than hardware.

9.2 Pricing and margins

Precision six-axis F/T transducers are low-volume, high-mix instruments. Publicly disclosed item-level prices are scarce. The note does not assert a specific ASP because no primary-source vendor price list was retrieved. Any specific price range would be needs primary source validation.

The inference from adjacent precision-instrument and robotics businesses is that gross margins are healthy but lower for integrated cobot volume than for laboratory-grade transducers. This inference is not sourced to a specific filing and is labeled as inference.

9.3 Who captures value

The sensor maker captures value only if the robot OEM/integrator cannot internalize the capability. The countervailing force is that large robot OEMs — especially humanoid developers with custom actuator programs — may design internal F/T sensing to control cost and integration. Tesla’s emphasis on in-house actuators and sensors, if realized at scale, would compress third-party sensor vendor value capture in that account [S8][S9]. By contrast, cobot OEMs often buy external F/T sensors because the transducer is specialized and the volumes do not yet justify in-house design.

Disconfirming evidence

The thesis that six-axis strain-gauge F/T sensing is a structural, growing value pool for humanoids and cobots would be weakened or invalidated by the following:

1. Volume humanoid production fails to materialize. If Tesla and other humanoid developers do not reach production volume, the specialized ankle/wrist F/T market remains a lab-scale niche [S9].
2. Cobot OEM internalization. If Universal Robots, Fanuc, Yaskawa, or others fully internalize F/T sensing and stop buying from third-party transducer vendors, independent sensor makers lose the cobot volume they currently serve [S10].
3. Alternative transduction displaces strain gauges. Capacitive, optical, magnetoelastic, or semiconductor MEMS sensors could deliver lower cost, lower drift, or easier integration at scale. The current company literature emphasizes strain-gauge-based six-axis sensors [S3][S4], but displacement risk is not disproven.
4. Force control loses share to vision/learning-only control. If high-bandwidth force control is replaced by vision-based manipulation and learned contact models, the marginal demand for precision F/T transducers would shrink [S2].
5. Calibration cost and fragility remain unsolved. Six-axis F/T sensors require calibration and can be damaged by overload. If these operational issues persist, adoption in high-volume humanoid fleets would be constrained.