BLOGS

Humanoid demos are sold on vision; humanoid deployments are decided by force. A production humanoid typically carries around 4 six-axis force/torque sensors and roughly 28 joint torque sensors — about 32 channels of force feedback that together account for a double-digit share of the bill of materials. This analysis walks the force-sensing stack layer by layer: where each sensor actually sits inside the joint, which measurement principle wins in which position, the engineering numbers that predict field behaviour rather than bench behaviour, and the specification checklist that decides whether a joint feels like a hand or like a hinge.

Vision Tells the Robot Where; Only Force Tells It What Is Happening

A camera can localize a bottle to within a millimetre. It cannot tell the robot whether the bottle is full or empty, whether the cap is cross-threaded, or whether the shelf behind it is about to tip. Every task that involves contact — inserting a connector, seating a bearing, wiping a surface, tightening a fastener, handing an object to a person — is governed by quantities that no optical sensor measures.

That is why the industry conversation has shifted. In the last two years the question posed to robot hardware teams has moved from *can it run and jump* to *can it actually do work*. Running and jumping is largely a kinematics and stability problem. Doing work is a contact problem, and contact problems are answered by torque.

• **Vision answers *where*.** Position, orientation, free-space planning, grasp-point selection.
• **Force answers *how much*.** Contact pressure, insertion force, slip onset, load distribution, collision severity.
• Neither substitutes for the other. Force sensing does not replace vision; it takes over the control authority the moment the robot touches something.

A useful mental model: vision is the robot's estimate of the world before contact, and force sensing is its measurement of the world during contact. Control quality during contact is what customers pay for, and it degrades the instant the force loop is missing or slow.

Two Feedback Loops, One Control Problem

Humanoid force sensing is usually described as one system. In practice it is two loops with very different physics and very different bandwidth requirements.

The proprioceptive loop measures the robot's own load state from the inside. Joint torque sensors, motor-side position encoders, and the torso IMU together tell the controller how much torque each joint is transmitting and where the body's centre of mass sits relative to the support polygon. This loop is what keeps a robot standing after an unexpected shove, and what lets it modulate ankle and hip torque as ground conditions change from concrete to gravel.

The exteroceptive loop measures interaction with the outside world. Six-axis force/torque sensors at the wrists and ankles report the resultant force and moment at those interfaces, while tactile arrays in the fingers report contact distribution and slip.

The distinction matters because of what each loop *cannot* see. A wrist-mounted force/torque sensor measures the resultant of everything below it, which means it cannot distinguish an object's weight from a collision with the forearm. A joint torque sensor measures the load path through its own joint, which is exactly what is needed for collision detection and hand-guiding along the link, but it says nothing about how a fingertip is loaded. Designing with one loop and claiming the other's functionality is one of the most common specification errors in humanoid hardware.

Where a Joint Torque Sensor Actually Sits

There is no single correct answer to "where does the torque sensor go", and the choice constrains everything downstream.

• Motor side (inferred from current). No added hardware, no added compliance. Also no direct measurement: friction, reducer losses and torque ripple all sit between the motor and the load, so the estimate degrades exactly where precision matters most.
• Between the reducer and the link. The classic arrangement for collaborative arms. The sensor sees the true output torque, including the load's own gravity term, which is what makes gravity compensation and hand-guiding feel natural. The price is a compliance element inserted in the structural chain, plus exposure to the non-torque loads — axial force, radial force, bending moment — that travel through the same interface.
• On the flexspline of the harmonic drive. Mechanically elegant, because the strain signal appears where the deformation already happens. Practically difficult: the flexspline deforms continuously under the wave generator at running speed, so the measured signal is modulated by speed and torque ripple, and routing strain-gauge wiring onto a rotating, flexing cup is a manufacturing problem before it is a sensing problem.
• Torsion-bar plus encoder. Torque is inferred from the twist of a calibrated bar. Less sensitive to reducer ripple, but the bar must be compliant enough to produce a measurable rotation, and that compliance is subtracted directly from arm accuracy.

The design equation is blunt: the torque sensor is a spring you deliberately added to a stiffness-critical structure. Every unit of sensitivity you buy with compliance is paid for in positioning accuracy, settling time and control bandwidth.

Why Motor Current Is Not a Torque Sensor

Current-based torque estimation is popular because it is free. On a geared joint with a high-ratio harmonic reducer, it is also the weakest of the available options, and it is worth being explicit about why.

The practical conclusion is that current sensing is an acceptable *supervisory* signal and a poor *control* signal. Systems that need safe hand-guiding, sub-newton contact resolution, or collision detection along the link generally pay for a real torque sensor at the joint.

The Five Numbers That Predict Field Behaviour

Datasheets often lead with range and accuracy. For humanoid joints, five other quantities decide whether the sensor survives the application.

1. Torsional stiffness. This is the number that quietly sets the robot's positioning accuracy. Published designs illustrate the spread: a taper-lock spoke topology reached 617.5 kN·m/rad of torsional stiffness with a ±150 N·m capacity and 0.07 N·m resolution, while an optimized shear-beam structure reported 503.44 kN·m/rad torsional and 142.56 kN·m/rad bending stiffness at a ±200 N·m range. Both designs exist specifically to avoid inserting a soft element into a stiff joint.

2. Crosstalk. Non-torque loads — axial force, radial force, bending moment, and the reaction of whatever tool is mounted downstream — leak into the torque reading. Uncompensated, that leak becomes torque error that the controller cannot distinguish from real load. Optimized structures have driven this down to the 0.74% level with in-structure cancellation, and to measurement errors below 0.5% of full scale under a 392.4 N axial load or a 600 N radial load, and under 1% of full scale under a 25 N·m bending moment — without resorting to an expensive cross-roller bearing. In the six-axis domain, structural plus algorithmic decoupling typically lands cross-talk in the 0.3%–0.5% FS band, with linearity around 0.3% FS.

3. Overload capacity. Humanoid joints see impacts, not just torques. Falls, stair missteps, collisions with fixtures and dropped payloads all produce transients far above rated torque. A sensor rated at 200%–300% of full scale for safe overload is not a luxury; it is the difference between a serviceable joint and a repair ticket.

4. Sensitivity. Structural optimization is the lever here, not amplifier gain — gain mainly matches signal amplitude to the ADC's dynamic range, while the elastic body decides how much strain the gauges actually see. Reported sensitivities across spoke topologies span an order of magnitude, from roughly 1.65 mV/N·m for an inclined sensing surface, through 17.48 mV/N·m for a taper-lock design and 25.71 mV/N·m for an optimized shear beam, up to 33.49 mV/N·m for a dumbbell-hole spoke and 38.67 mV/N·m for a pear-shaped-hole design.

5. Resolution, linearity and hysteresis. These determine the smallest force the robot can reliably resolve and whether it returns to the same reading after unloading. Comparable designs report linearity near 0.2%, repeatability around 0.06%, and hysteresis near 0.5%. Hysteresis is the one that hurts in force control: it appears as a direction-dependent error that no amount of calibration removes.

The trade-off is structural, not incidental. Sensitivity rises as the elastic body gets softer, and softness is subtracted from joint stiffness. Multi-objective design — optimizing sensitivity, stiffness and crosstalk together rather than in sequence — is the only route to a sensor that improves one without silently degrading the other two.

Six-Axis Sensors: A Thin-Profile Problem

A six-axis force/torque sensor resolves three orthogonal forces and three moments simultaneously. On a humanoid, the demanding positions are the wrist and the ankle, and the binding constraint there is not range but geometry.

• Thickness is the specification. Wrist and ankle envelopes are measured in millimetres, and a sensor that is 10 mm thick is already a serious packaging achievement. Anything thicker pushes the hand further from the arm axis or the foot further from the ground, changing the very dynamics the sensor is meant to measure.
• Hollow bores matter as much as range. Cable and tendon routing has to pass through the joint. A sensor with a usable central bore lets the harness travel along the kinematic axis instead of looping around it, which is often the difference between a serviceable limb and one that fails at a flex point.
• Foot contact is a special case. The ankle sensor doubles as the ground-reaction estimator, which makes its overload behaviour a safety property rather than a performance metric. Foot-strike transients are short, large and frequent.
• Cross-axis accuracy is the real deliverable. What customers experience as "the robot is shaky when it pushes" is frequently crosstalk rather than control tuning.

Fingertip Tactile: Where Form Factor Beats Range

At the fingertip the priority order inverts. Range is small, thickness is microscopic, and the value of the sensor is entirely in resolution and update rate.

Micro force/torque units in the 6.5 mm to 9.5 mm class now fit at the fingertip and between finger phalanges. Joint torque sensors have followed the same miniaturization curve, with diameters down to the 10–14 mm band covering different joint envelopes and load classes. Two measurement routes compete: MEMS-based elements, which scale well and integrate readout close to the sensing site, and foil strain-gauge elements, which deliver very fine resolution — foil gauges are routinely specified to resolve deformation below one microstrain, with product ranges spanning tens of newtons on a single axis up to kilonewton-class six-axis units.

What is still unsolved is not the sensing element but the system around it. Durability under millions of contact cycles, the sheer data volume of high-density arrays, and the absence of a common sensor interface across vendors are the open problems the community itself flags. For integrators, that means tactile channels should be treated as a data pipeline with its own bandwidth and protocol budget, not as an accessory bolted to the end effector.

The Joint Count Is the Business Case

The structural reason force sensing crossed from optional to standard is arithmetic, not fashion.

• A traditional industrial arm carries on the order of 7 joints. A humanoid can exceed 40 degrees of freedom.
• Industry estimates for a production humanoid cluster around 4 six-axis units at wrists and ankles plus roughly 28 joint torque sensors distributed through the body — on the order of 32 force channels per robot, with the exact split dependent on architecture and use case.
• Those channels together represent a double-digit percentage of the bill of materials, a share that is large enough to matter in robot economics and small enough that removing it is not a real option.

This changes the supplier problem. A component that appears four times per robot can be built in a low-volume, high-touch workshop. A component that appears twenty-eight times per robot has to be produced on an automated line, calibrated in a repeatable process, and delivered with batch consistency — because a joint-to-joint mismatch in torque calibration shows up in the field as asymmetric gait, not as a failed part. It is the same shift that precision reducers went through as they moved from industrial arms to humanoid joints, and it is now happening to the sensing layer.

Integration: Stiffness, Bandwidth and the Control Loop

A torque sensor only improves a robot if the rest of the control chain can use it. Three integration questions dominate.

Bandwidth matching. Current loops in modern joint drives close in the tens of kilohertz. Force loops close far lower — tens of hertz for impedance-style interaction control. If the sensor's effective bandwidth sits below the force loop's crossover, the sensor becomes the limiting pole: phase lag accumulates, gain margins shrink, and the joint oscillates exactly when the operator is pushing on it. Specifying a fast sensor and a slow signal chain produces the same bad outcome as specifying a slow one.

Synchronisation and filtering. Thirty-two force channels sampled asynchronously will produce a body-load estimate that is wrong by however much time-skew separates them. Simultaneous sampling and explicit anti-alias filtering are control-system requirements, not data-acquisition niceties.

Thermal behaviour. Joint interiors near motors get hot; sensors in that region will drift. Components rated for operation up to roughly 105 °C exist precisely for this, but a temperature rating is not the same as thermal compensation. Zero-point drift with temperature is a systematic error that shows up as a slowly changing force bias, and it must be compensated in software or in hardware — it will not average out.

The payoff of getting this right is the behaviour users describe as "the robot knows how hard it is pushing": compliant insertion without force spikes, grasp force that adapts to whether the object is an egg or a wrench, and a body-load estimate stable enough to keep the robot upright on uneven ground.

The Wiring, Calibration and Service Tax

Three costs are routinely underestimated at the specification stage.

1. Routing. Every force channel needs power, signal and, in many cases, a shield or a digital bus. Inside a hollow-shaft joint with a limited bend radius, harness design becomes a mechanical engineering problem with a real fatigue life.
2. Calibration at scale. Per-unit calibration is unavoidable; the question is whether it is manual and slow or automated and traceable. With dozens of channels per robot, calibration throughput is a production constraint on the supplier and a line constraint on the integrator.
3. Serviceability. Sensors live in the load path and will eventually be replaced. A design that requires de-stacking the joint to swap a sensing element converts a routine maintenance action into a rebuild. Designing the sensor as a serviceable module rather than an embedded component is an architectural decision made at the beginning or not at all.

What to Specify Before You Sign Off

• Placement: output-side torque sensing for interaction control; confirm the compliance penalty against the joint's positioning budget.
• Crosstalk under worst-case non-torque load: ask for measured error under axial, radial and bending loads, not just nominal torque accuracy.
• Safe overload: target at least 200%–300% of full scale, validated against the actual impact case, not a generic shock value.
• Stiffness: torsional stiffness in kN·m/rad, compared against the reducer's own stiffness so the sensor does not become the softest element in the chain.
• Hysteresis and repeatability: these are what force control actually feels like; they are more useful than headline accuracy alone.
• Bandwidth and sampling: sensor and signal chain bandwidth compared against the intended force-loop crossover, with synchronised sampling across all channels.
• Thermal drift: specified zero-point drift over the operating temperature range, and whether compensation is in the sensor or in the controller.
• Mechanical integration: hollow bore diameter, harness path, bend radius, and whether the sensor can be replaced without dismantling the joint.
• Channel economics: calibration traceability and batch consistency across the dozens of units one robot consumes — this is where a sensor programme succeeds or fails at scale.

The Ti5ROBOT Engineering View

Force sensing is only half of the sensing problem; the other half is whether the joint can *act* on the signal. A torque reading is useful to a controller only if it arrives on time, with known latency, and if the drive can modulate current accurately enough to respond within the force loop's settling window. That is why integrated joint design treats sensing, control compute and power electronics as one system rather than three procured parts.

Two implications follow for anyone specifying humanoid joints:

• The sensor is only as good as the loop it sits in. Choose the sensing layer against the control bandwidth you intend to run, then confirm that the joint's drive electronics, encoder feedback and compute determinism can actually sustain that loop. Our Lightweight Joint Module platform and the more detailed 30–40 mm joint module family are built around that coupling, with the electrical and mechanical integration decided together rather than bolted together.
• Sensing architecture belongs in the joint choice, not after it. Deciding where torque is measured changes the compliance budget, the harness routing, the thermal path and the achievable force bandwidth. It is a first-order decision. The Humanoid Robot Actuator platform is specified on the same premise: the joint is a control system with a mechanical output, not a gearbox with a motor attached.

Two earlier analyses frame the rest of the picture. Inside the Integrated Joint covers the compute and encoder side of the same accuracy ceiling, and One Humanoid, Three Reducers maps how transmission choice varies joint by joint across the body. Read together with this piece, the three describe the same design object from three directions: what moves the load, what measures the load, and what decides what to do about it.

Source: Adapted and independently rewritten from a September–October 2026 Chinese-language industry survey of robot force-sensing supply chains, supplemented with peer-reviewed joint-torque-sensor design literature (strain-gauge spoke topologies, crosstalk and stiffness optimisation) and English-language force/torque sensor reporting. Sensor-count and bill-of-materials share figures reflect industry estimates and vary with robot architecture and use case; peer-reviewed stiffness, crosstalk, sensitivity and resolution values are design-specific and quoted as published.