One joint, three sensors

A humanoid joint module packs three sensing jobs into a housing the size of a fist. The motor-side encoder sits on the rotor and runs fast: it commutates the permanent-magnet motor and feeds the velocity loop, so update rate matters more than absolute accuracy — incremental optical or magnetic devices with 10,000-plus counts per revolution are typical. The output-side encoder sits after the gearbox and answers a different question: where is the limb, right now, including at power-on? That calls for an absolute encoder, because a robot that has to wiggle every joint through a homing routine before it can stand up is not a product.

Why not just divide the motor angle by the gear ratio? Because the transmission is not rigid. Under load, a harmonic drive’s flexspline winds up, gear teeth deflect and backlash opens, so the motor’s idea of the joint angle and the limb’s actual angle disagree — by exactly the amount the transmission is straining. A single motor-side encoder cannot see that error. Dual encoding fixes it, and as a bonus turns the transmission into a spring scale: if the stiffness k of the drivetrain is known, torque can be estimated as τ = k(θ_motor/N − θ_output). The third sensing job, torque, then has four competing answers, which is where the architectures diverge.

Magnetic, optical, capacitive, inductive

Four encoder technologies compete for those two slots per joint, and none wins everywhere. Magnetic encoders dominate the output side today because they are non-contact, tolerant of dust and shock, and forgiving of installation error. RLS, the Slovenian encoder maker owned by Renishaw, markets a three-product stack straight at humanoid builders: the AksIM-2 off-axis ring (up to 20-bit, ring diameters from 29 to 80 mm) for shoulders, hips, elbows and knees; the through-shaft Orbis (up to 14-bit) where cabling has to pass through the joint; and the 8 mm RM08 (up to 12-bit) for fingers and grippers. PAL Robotics has integrated RLS magnetic encoders in its humanoids for balance-critical joints, and Renishaw has built custom AksIM variants for TQ-RoboDrive’s ILM robot-joint motors under a 28 mm outer-diameter constraint.

The catch is that a magnetic encoder’s worst enemy — a stray magnetic field — is exactly what a frameless permanent-magnet motor produces a few millimetres away. That is the opening for the alternatives. Capacitive encoders reach around 19-bit resolution with update rates above 10 kHz and are immune to magnetic interference. Inductive encoders, such as Celera Motion’s IncOder line and FLUX’s rotary encoders, make the same immunity claim while staying lightweight and hollow-shafted, and are increasingly pitched at legged and surgical robots. Optical incremental encoders still offer the highest counts and sub-microsecond latency, but lose absolute position at power-off and dislike dust and thermal cycling — which keeps them on the motor side, if they are used at all.

Four ways to feel torque

Position sensing has converged; torque sensing has not. The cheapest method is to not have a sensor at all: in a quasi-direct-drive joint with a gear ratio below about 10:1, motor current is a usable proxy for output torque, with control resolution around 1–2 percent of rated torque. The physics only works at low ratios. Push the same trick through a 120:1 harmonic drive and a newton-metre of contact at the output shows up as roughly 8 mN·m at the motor — buried under friction that the ratio amplifies by four orders of magnitude. That is why current-based sensing belongs to the QDD camp (CubeMars’ AKE80-8 runs 8:1; RobStride’s 17 N·m modules pair the low ratio with dual encoders), and why high-ratio joints need hardware.

The hardware options are three. Deflection-based sensing reuses the dual encoders: measure the windup across a compliant element of known stiffness — a series elastic actuator makes the spring explicit, a harmonic drive’s flexspline provides one for free — and compute torque from the angle difference. Dedicated strain-gauge torque sensors bond gauges to a machined flexure between gearbox and limb and resolve better than 0.1 N·m, at the price of roughly 30 percent of a rotary actuator’s bill of materials, per HONPINE’s published cost breakdown. And at the wrists and ankles, where the robot meets the world, six-axis force/torque sensors from suppliers such as Bota Systems, FUTEK, ATI and Robotous measure all three forces and all three moments at once — market research puts their adoption at about 70 percent of new humanoid platforms, at five to eight times the price of a single-axis device.

What the robots actually ship

Tesla put its cards on the table at AI Day 2022: the rotary Optimus actuator slide listed dual encoders and an integrated torque sensor alongside the harmonic drive, and the linear units carry force sensors in series with the rod. Reconstructions of the current design describe every structural actuator as an integrated module — motor, reducer, encoder and driver board pre-assembled and calibrated as a unit. The QDD camp goes the other way: Unitree-style joints and the commercial modules from CubeMars and RobStride skip the torque sensor entirely and lean on low ratios plus current control, which is a large part of why their actuators undercut Western designs on price. In between sit the compliance-first designs — Apptronik has said publicly that it tested 30 actuator generations before settling on Apollo’s force-controlled linear architecture.

The supplier map is redrawing itself around that split. On encoders, RLS/Renishaw, Netzer, Celera Motion and FLUX are joined by the chip-level players whose magnetic sensor ICs sit inside cheaper integrated modules. On torque, the specialist force/torque houses — Bota Systems, ATI, FUTEK, Robotous — now compete with Chinese sensor makers such as Keli scaling for humanoid volume. Intel Market Research sizes torque sensors for humanoids at roughly $600 million in 2025, growing to a projected $10.5 billion by 2034 at a 44.3 percent CAGR, with Asia-Pacific already the largest supplying region. Even discounting forecast optimism, the direction is clear: sensing is where a meaningful share of the joint’s future bill of materials is moving.

FAQ

Sources