Force-torque sensor

A force-torque sensor (F/T sensor) is a transducer that measures three orthogonal forces (Fx, Fy, Fz) and three torques or moments (Tx, Ty, Tz) applied simultaneously at a single point, across six degrees of freedom, most often at a robot's wrist, ankle, or tool flange. The dominant commercial design bonds strain gauges to a compliant metal structure, frequently a spoked or "Maltese cross" beam array, that deforms by only microns under load; capacitive, piezoelectric, and optical designs trade some of that precision for cost, ruggedness, or bandwidth. In humanoid and industrial robotics, F/T sensors are the standard way to give a machine a quantitative sense of contact, enabling force and impedance control, compliant assembly, and collision detection, but a growing number of humanoid-robot teams instead infer joint load from motor current and encoder position, an approach that cuts cost and weight at the expense of some accuracy.

How force-torque sensors work

A six-axis F/T sensor is built around two rigid mounting plates joined by a compliant metal element. One plate typically bolts to the last link of a robot arm; the other bolts to a tool, gripper, or other end effector, or to the ground contact point on a foot. Between the plates, a machined structure, commonly aluminum, stainless steel, or titanium, is cut into a spoked or cruciform shape that the engineering literature calls a "Maltese cross" beam array ^[1]. Under load, the spokes flex by only a few microns, well inside the metal's elastic limit, so the structure springs back to its original shape once the load is removed ^[1].

Strain gauges, thin metal-foil or semiconductor elements whose electrical resistance changes slightly when stretched or compressed, are bonded to the spokes at points chosen so each gauge responds strongly to some load directions and weakly to others ^[2]. The gauges are wired into Wheatstone bridge circuits, four-element networks that turn a tiny resistance change into a proportional voltage; a typical industrial sensor carries two dozen or more gauges arranged into six or more bridges ^[3].

Because the spokes are mechanically connected, no single bridge responds to only one axis of loading; pushing straight down also strains the gauges meant to sense sideways force, an effect the industry calls crosstalk. Manufacturers correct for this in two stages. Raw bridge voltages are first amplified and digitized, then multiplied by a calibration, or decoupling, matrix, a 6-by-6 matrix determined by loading a finished sensor with known reference forces on a calibration rig, to convert the six coupled channel readings into clean Fx, Fy, Fz, Tx, Ty, Tz outputs ^[3][1]. Some designs also reduce crosstalk mechanically, before any correction is applied, by optimizing the taper and cross-section of the spokes themselves ^[1].

In brief: the sensor is a small, deliberately flexible metal skeleton wearing a dozen or more strain gauges. The robot's controller never reads raw strain values; it reads six pre-translated numbers, three pushes and three twists, computed against a factory calibration table specific to that individual unit.

Overload is the sensor's main failure mode. If an applied force exceeds the rated range, the elastic structure can deform permanently, which invalidates the calibration matrix and can crack the gauges themselves. Commercial units are typically rated with overload margins of roughly five to twenty times their nominal sensing range, and vendors sell shock-tolerant variants aimed at humanoid and mobile-manipulation use, where unplanned impacts are routine rather than exceptional ^[4].

Transduction technologies

Four transduction principles dominate commercial and research designs, each with a different tradeoff between accuracy, cost, and durability.

Scroll sideways for more →

Despite the newer alternatives, commercially available six-axis sensors are still dominated by piezoresistive strain-gauge designs; capacitive and optical approaches are gaining ground mainly where their tolerance to overload and impact, or their potential for low-cost mass production, outweighs strain gauges' edge in raw accuracy ^[14].

Calibration, crosstalk, and key evaluation criteria

Integrators generally evaluate an F/T sensor against a fairly consistent checklist: full-scale range on each of the six channels, resolution (the smallest detectable change), single-axis crosstalk error (often specified as a percent of full scale), sample rate and communication interface (analog, CAN, EtherCAT, or Ethernet), overload rating, physical size and mass, and environmental sealing rating ^[5][4]. For a humanoid robot, mass and diameter at the wrist or ankle matter disproportionately, since the sensor sits at the far end of a limb, where added weight increases the torque the rest of the arm or leg has to work against ^[14].

Calibration itself degrades over a sensor's service life. Temperature swings, mechanical fatigue, and, for strain-gauge units, gauge-bond aging can all shift the factory decoupling matrix away from the values it shipped with. Researchers have proposed in-situ recalibration methods that let a robot re-derive its own sensor's calibration from known contact geometry, rather than shipping the unit back to a calibration lab, an approach aimed specifically at keeping humanoid robots accurate over months of field use ^[15].

Dedicated sensors versus sensorless force estimation

Whether a robot needs a dedicated F/T sensor at all is an active, unresolved question in humanoid-robot design, not one with a single settled answer.

The case against a dedicated sensor is straightforward. An industrial-grade six-axis unit adds cost, ranging from roughly a hundred dollars for the cheapest research-grade parts to well over a thousand dollars per unit for calibrated industrial models, mass at the worst possible location on a limb, an extra wiring run, and a component that can drift out of calibration or fail on impact. A humanoid needing sensors at both wrists and both ankles multiplies those costs before a single finger is counted ^[14].

The oldest engineering answer to that tradeoff was not to remove force sensing but to move it inside the joint. Starting in the 1980s and maturing through the German Aerospace Center's (DLR) lightweight-robot program in the early 2000s, researchers led by Gerhard Hirzinger and Alin Albu-Schaffer sensorized the flex spline of each joint's harmonic drive gear directly, measuring joint torque at the actuator rather than only at the tool flange. That design enabled active vibration damping and whole-arm impedance control without any wrist-mounted sensor, and its influence carries through to many of today's torque-controlled robot-arm joints ^[16].

A newer, cheaper alternative has grown up alongside quasi-direct-drive actuators: legged and humanoid platforms built around low-reduction-ratio joints, where gearbox friction is small enough that motor torque stays roughly proportional to motor current. Because the motor driver already measures current to commutate the motor, the robot gets an approximate joint-torque reading with no added hardware, using roughly the relationship torque equals a motor torque constant times current ^[17]. Published technical descriptions of quasi-direct-drive legged and humanoid platforms, including Unitree's G1, describe estimating joint torque and ground contact this way rather than through a dedicated torque sensor ^[18]. The tradeoff is measurable: one comparison of joint torque sensors against motor-current-based estimates on a legged robot found the dedicated sensor far more accurate (R-squared of 0.9998, root-mean-square error of 0.032 newton-meters) than the current-based estimate (R-squared of 0.961, root-mean-square error of 0.164 newton-meters) ^[19]. Highly geared actuators behave differently: friction and backlash inside a high-ratio gearbox corrupt the current-to-torque relationship enough that current sensing alone is considered unreliable at low speed, which is part of why designs built around such actuators have historically paired them with dedicated strain-gauge torque sensors instead ^[16][19].

A third path, still mostly confined to research, tries to close that accuracy gap through software rather than hardware. Momentum-observer methods and, more recently, learning-based estimators, including physics-informed neural networks combined with Kalman filtering and limb-modularized networks trained on joint encoder and inertial data, aim to infer external torque and contact force purely from proprioception (position, velocity, and current), without any force or torque transducer in the loop ^[20][21][22].

Flexiv, a Shanghai-headquartered robotics company founded in 2016 by a team from Stanford University's robotics labs, is a prominent example pulling in the opposite direction from sensorless designs ^[23]. Rather than removing force sensing, Flexiv's Rizon arms build proprietary force- and torque-sensing technology into every joint, not only the wrist; the company states this enables "direct force control" and force-sensing accuracy down to 0.03 newtons ^[24]. Flexiv describes the resulting sensing density as giving its arms "muscle-like" responsiveness to contact, a company characterization rather than an independently verified figure, and points to contact-rich applications such as handling delicate fish fillets on a moving conveyor and automated massage therapy as places where dense force feedback outperforms position-only control ^[25].

Neither approach has won outright. Current-based estimation on quasi-direct-drive joints is well suited to whole-body balance and coarse collision response, where approximate torque is good enough and every gram of limb mass matters for dynamics. Dense, dedicated force sensing remains attractive wherever a task needs fine, repeatable force control, such as precision assembly or handling fragile objects, where its added cost and weight are easier to justify.

Use in humanoid robots

Compliant control and assembly. A wrist-mounted F/T sensor lets a controller run impedance or admittance control, in which the arm behaves like a programmable spring and damper rather than a rigid position-following mechanism, "giving" when it meets unexpected resistance instead of fighting it ^[2][14]. This underlies hybrid force and position control strategies used for tasks like peg-in-hole assembly, where the arm searches for a hole, establishes contact, and inserts a part guided by real-time force feedback rather than a pre-programmed blind trajectory ^[26].

Collision detection and safety. A sudden, unplanned change in measured force, from a human touching the arm or the arm striking an obstacle, can be detected within milliseconds and used to trigger a stop or a compliant retreat ^[2]. Power and force limiting is one of the recognized modes of safe operation for a collaborative robot. As of a 2025 revision that took effect that April, the collaborative-application safety requirements formerly published as the standalone ISO/TS 15066 technical specification were folded directly into the main ISO 10218 industrial-robot safety standard ^[27][28].

Balance in bipedal locomotion. Many legged and humanoid robots mount a six-axis F/T sensor, or an array of load cells configured to produce equivalent six-axis output, at or near each ankle to measure the ground reaction force and moment. That measurement feeds the zero moment point calculation widely used to judge whether a walking robot is about to tip over and to correct its gait in real time ^[29][30].

Hands and fingertips. Full six-axis F/T sensors are generally too large and heavy to fit inside a dexterous hand's fingertip, so fine-grained contact sensing there is normally handled instead by distributed tactile sensing arrays or electronic-skin-style pads, with a single wrist-level F/T sensor providing the coarser, whole-hand force reading ^[14].

Across humanoid robot platforms, sensor-vendor and academic literature describes force and torque sensing being applied at wrists, ankles, and, in some designs, other joints to support robot manipulation and locomotion, though individual manufacturers vary widely in how much of their sensor architecture they disclose publicly ^[14][18].

Suppliers and landscape

The F/T sensor supply chain has two distinct layers that are easy to conflate. A handful of specialist manufacturers, including ATI, Bota Systems, OnRobot, Kistler, and Schunk, design, machine, calibrate, and sell complete, ready-to-mount six-axis sensor products. Behind them sit component-level suppliers of strain gauges, load cells, and sensing elements, such as Sensata Technologies and TE Connectivity, whose piezoresistive strain-gauge and load-cell technology is more often designed into force-measurement products across automotive, aerospace, and industrial markets, including robot joints and end effectors, than sold directly as a branded six-axis wrist sensor ^[31][32]. Treating the two layers as interchangeable, as a flat supplier list can imply, overstates how many companies actually compete head-to-head on complete F/T sensor products.

Scroll sideways for more →

Market-research estimates for the six-axis F/T sensor market vary widely and should be read as directional rather than precise, given how new and fast-changing humanoid-robot demand is. QYResearch put the global market at roughly 375 million dollars in 2025, projecting growth to about 4.6 billion dollars by 2032 at a 43.7 percent compound annual rate; a separate estimate from Valuates Reports put 2025 revenue closer to 400 million dollars, growing to roughly 4.5 billion dollars by 2032 ^[41][42]. Both firms cite humanoid-robot adoption as the primary growth driver, but neither figure is an audited or independently verified number, and estimates from market-research vendors in this space have historically diverged by wide margins.

See also

• Actuator
• Tactile sensing
• Dexterous hand
• Robot manipulation
• Humanoid robot
• Quasi-direct drive
• Rotary encoder
• Harmonic drive
• Flexiv

References

1. "Improving the coupling errors of a Maltese cross-beams type six-axis force/moment sensor using numerical shape-optimization technique," Mechanism and Machine Theory, ScienceDirect. https://www.sciencedirect.com/science/article/abs/pii/S0263224118304640 ↩
2. Robotiq, "Robotics Research 101: How Do Force Sensors Work?" blog.robotiq.com. https://blog.robotiq.com/robotics-research-101-how-do-force-sensors-work ↩
3. "Decoupled Six-Axis Force-Moment Sensor with a Novel Strain Gauge Arrangement and Error Reduction Techniques," Sensors, 2019, via PMC. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6651882/ ↩
4. ATI Industrial Automation, Axia80 Low-Cost Force/Torque Sensor datasheet, DirectIndustry. https://pdf.directindustry.com/pdf/ati-industrial-automation/axia80-low-cost-force-torque-sensor/24443-712672.html ↩
5. ATI Industrial Automation (Novanta), "Force/Torque Sensors," ati.novanta.com. https://ati.novanta.com/products/force-torque-sensors/ ↩
6. Schunk, "Force/torque sensors," schunk.com. https://schunk.com/us/en/automation-technology/force/torque-sensors/c/PUB_11579 ↩
7. Kistler, "6-axis force-torque sensors" (9306A series), kistler.com. https://www.kistler.com/INT/en/cp/6-axis-force-torque-sensors-9306a/P0000699 ↩
8. Kistler, "Milestones," kistler.com. https://www.kistler.com/INT/en/milestones/C00000002 ↩
9. Stanford TechFinder, "Using capacitive sensing to create robust, low-cost force-torque sensors for real-world robotics," techfinder.stanford.edu. https://techfinder.stanford.edu/technology/using-capacitive-sensing-create-robust-low-cost-force-torque-sensors-real-world-robotics ↩
10. "Miniaturized silicon-based capacitive six-axis force/torque sensor with large range, high sensitivity, and low crosstalk," Microsystems and Nanoengineering, Nature, 2024. https://www.nature.com/articles/s41378-024-00831-0 ↩
11. OnRobot, "Hex 6-Axis Force Torque Sensor," onrobot.com. https://onrobot.com/en/products/hex-6-axis-force-torque-sensor ↩
12. Fierce Sensors, "Robotic Apps Benefit From Gripper And Force-Torque Sensor Upgrades." https://www.fiercesensors.com/components/robotic-apps-benefit-from-gripper-and-force-torque-sensor-upgrades ↩
13. "Self-uncoupling wrist six-axis force/torque sensor based on Fiber Bragg grating," Mechanism and Machine Theory, ScienceDirect. https://www.sciencedirect.com/science/article/abs/pii/S0263224123003913 ↩
14. "Progress in multi-axis force sensors: Sensing mechanisms, structural design, and applications in humanoid robots," Sensors and Actuators A: Physical, ScienceDirect, 2026. https://www.sciencedirect.com/science/article/abs/pii/S0924424726006138 ↩
15. "Nonlinear In-situ Calibration of Strain-Gauge Force/Torque Sensors for Humanoid Robots," arXiv:2312.09846. https://arxiv.org/pdf/2312.09846 ↩
16. G. Hirzinger, N. Sporer, A. Albu-Schaffer, et al., "On a new generation of torque controlled light-weight robots," Proceedings of the IEEE International Conference on Robotics and Automation, 2001. https://www.diag.uniroma1.it/deluca/pHRI_elective/ICRA01_New_generation_light-weight_robots_Hirzinger_etal.pdf ↩
17. "Application of Proprioception Quasi-Direct Drive Actuators," eScholarship, University of California. https://escholarship.org/content/qt9dk9r8d3/qt9dk9r8d3_noSplash_d13e8322e709c1d65d7a255b8c6a7404.pdf ↩
18. Unitree G1 technical overview, docs.quadruped.de, and secondary technical coverage of quasi-direct-drive humanoid actuator architecture. https://docs.quadruped.de/projects/g1/html/g1_overview.html ↩
19. "Contact Sensing via Joint Torque Sensors and a Force/Torque Sensor for Legged Robots," arXiv:2510.10843. https://arxiv.org/pdf/2510.10843 ↩
20. "Physics-Informed Neural Networks with Unscented Kalman Filter for Sensorless Joint Torque Estimation in Humanoid Robots," arXiv:2507.10105. https://arxiv.org/html/2507.10105v1 ↩
21. "MOB-Net: Limb-modularized Uncertainty Torque Learning of Humanoids for Sensorless External Torque Estimation," arXiv:2402.11221. https://arxiv.org/html/2402.11221 ↩
22. "Proprioceptive External Torque Learning for Floating Base Robot and its Applications to Humanoid Locomotion," arXiv:2309.04138. https://arxiv.org/abs/2309.04138 ↩
23. "Flexiv," Wikipedia. https://en.wikipedia.org/wiki/Flexiv ↩
24. Flexiv, Rizon product page. https://www.flexiv.us/products/rizon ↩
25. Flexiv, case studies. https://www.flexiv.com/case-studies ↩
26. "Research on Robotic Peg-in-Hole Assembly Method Based on Variable Admittance," Applied Sciences (MDPI), 2025. https://www.mdpi.com/2076-3417/15/4/2143 ↩
27. "Evolution of safety requirements in industrial robotics: Comparative analysis of ISO 10218-1/2 (2011 vs. 2025) and integration of ISO/TS 15066," ScienceDirect. https://www.sciencedirect.com/science/article/pii/S2590123026015203 ↩
28. Association for Advancing Automation (A3), "Updated ISO 10218: Frequently Asked Questions." https://www.automate.org/robotics/blogs/updated-iso-10218-faq ↩
29. "Multi-Axis Force-Torque Sensors for Measuring Zero-Moment Point in Humanoid Robots: A Review," IEEE Xplore. https://ieeexplore.ieee.org/document/8871222/ ↩
30. "Development of a Six-Axis Force and Torque Sensor for the Humanoid Robot Sweaty 2.0," ResearchGate. https://www.researchgate.net/publication/313497201_Development_of_a_Six-Axis_Force_and_Torque_Sensor_for_the_Humanoid_Robot_Sweaty_20 ↩
31. Sensata Technologies, "Force Sensors," sensata.com. https://www.sensata.com/products/force-sensors ↩
32. TE Connectivity, "Load Cell Technologies," te.com. https://www.te.com/en/products/sensors/force-sensors/resources/load-cell-technologies.html ↩
33. The Robot Report, "Novanta acquiring ATI Industrial Automation for $172M." https://www.therobotreport.com/novanta-acquiring-ati-industrial-automation-for-172m/ ↩
34. Novanta Investor Relations, "Novanta Announces Agreement to Acquire ATI." https://investors.novanta.com/news/news-details/2021/Novanta-Announces-Agreement-to-Acquire-ATI/default.aspx ↩
35. Bota Systems, "About." https://botasys.com/about/ ↩
36. ETH Zurich Robotic Systems Lab, "Bota Systems spinoff." https://rsl.ethz.ch/partnership/spinoff/botasystems.html ↩
37. Schunk SE and Co. KG, company profile, wlw.com. https://www.wlw.com/en/company/schunk-se-co-kg-325839 ↩
38. Futu News, "Keli Sensing Technology (603662.SH): Six-dimensional force/torque Sensors have independently completed the series development of products for humanoid robot wrists, ankles, industrial arms, and collaborative arm ends." https://news.futunn.com/en/post/51666084/keli-sensing-technology-603662-sh-six-dimensional-force-torque-sensors ↩
39. Novanta Inc., corporate site and company profile. https://novanta.com/ ; https://stockanalysis.com/stocks/novt/company/ ↩
40. DCFModeling, "Sensata Technologies Holding plc: history, ownership, mission." https://www.dcfmodeling.com/blogs/history/st-history-mission-ownership ↩
41. QYResearch, "Robot Six-Axis Force Sensor" market report. https://www.qyresearch.com/reports/3520538/robot-six-axis-force-sensor ↩
42. Valuates Reports, "Global 6 Axis Force Sensors Market." https://reports.valuates.com/market-reports/QYRE-Auto-11O18850/global-6-axis-force-sensors ↩
43. "Novanta History," Pestel-analysis.com, citing company and SEC filings on the General Scanning Inc. to GSI Lumonics to GSI Group to Novanta corporate lineage. https://pestel-analysis.com/blogs/brief-history/novanta ↩