How Humanoid Joints Are Tested: Dynamometers, End-of-Line Checks and the Standards Still Being Written
How humanoid joints are tested: dynamometers, end-of-line checks and the standards still being written
A joint actuator’s datasheet is a set of claims. The test bench is where they get checked: torque against a reference sensor, slack and spring under load, efficiency across the whole operating map, heat after an hour of work. For humanoids, which need dozens of joints per robot, that bench is turning into a production bottleneck of its own.
Sources: Econotests, robot joint module test system; MIIT humanoid standards framework via eWeek. Figures for cobot-class joint modules; humanoid joints are tested with the same methods.
Why a joint has to prove itself before it goes in a robot
A humanoid has 20 to 40 actuated joints. One weak unit is enough to ground the whole machine, and you can’t find it once it’s buried inside a hip. So every serious joint maker runs two kinds of testing. Design validation takes days and characterises a new actuator model in depth. End-of-line testing takes minutes and checks that each unit coming off the assembly line matches the model it claims to be.
The methods aren’t new. Cobot makers have tested joint modules like this for a decade, and the test-equipment vendors – Magtrol, HBK, Kistler and a crowd of Chinese bench builders – sell the same hardware to e-motor and gearbox lines. What’s new is the volume humanoid OEMs are planning for, and the fact that nobody has yet agreed on a common way to report the results.
The bench: what actually gets bolted together
Strip away the enclosure and a joint test bench is a straight line of rotating parts. The joint under test sits at one end. A coupling connects its output to an in-line torque transducer, which also reads speed and angle. A second coupling connects the transducer to a load motor that can brake the joint, drive it backwards, or hold it still. Around that line sit a power analyzer on the joint’s DC supply, temperature probes on the winding and reducer housing, and a data logger tying all of it together on one time base.
Figure 1 · Animated · Layout of a joint actuator test bench
Illustration: humanoid.guide, simplified. Based on the test configurations described by Econotests, Lanmec and Teragon Research. The stripes moving across the shafts show them turning; animation is CSS-only and stops when reduced motion is requested.
The load motor is the clever bit. Because it can absorb power or supply it, the same bench can run the joint as a motor (lifting a leg) and as a brake (lowering the robot into a squat), which is exactly the mix a humanoid knee sees all day. Magtrol’s M-TEST 7 software, sold for robot motor testing, lists the routines this enables: ramp, curve, coast, pass/fail and overload-to-trip tests.
Eight measurements that make up a joint’s report card
Test-system maker Econotests breaks joint-module testing into eight core items. The acceptance values below are for collaborative-robot joints, which are the closest mature reference; humanoid joints built around quasi-direct-drive planetaries will usually land looser on backlash and better on bandwidth.
Table 1 · Core joint-module tests and typical acceptance values (cobot-class joints)
| Test | What it catches | Typical value or condition |
|---|---|---|
| Output torque accuracy | Gap between commanded and delivered torque | ±1–3% FS (±0.5% precision) |
| Torque ripple and cogging | Jerky motion at low speed | Cogging <0.5–2% of rated |
| Torsional stiffness | Wind-up under load, hysteresis | Measured in N·m/rad across the load range |
| Backlash | Slack on reversal | <1–3 arcmin at ±10% rated torque |
| Efficiency map | Losses across the torque-speed space | 65–80% at rated point |
| Thermal run | Derating as winding and reducer heat up | 60–120 min at rated load |
| Holding torque | Drift under a sustained gravity load | 30–60 min static |
| Dynamic response | Torque bandwidth (−3 dB point) | 10–50 Hz |
Source: Econotests, Robot Joint Module Test System: Torque, Stiffness and Efficiency for Cobot Actuators. FS = full scale. Values are typical industry ranges for cobot joints, not humanoid-specific requirements.
Bench builders add the functional checks on top. Lanmec’s 1,000 N·m joint test bench, for example, also verifies the holding brake, the overload and overcurrent protection, and communication stability on EtherCAT or CANopen – because a joint that measures perfectly but drops off the bus mid-stride is still a failed joint. Its durability programme runs above rated conditions to accelerate wear, with repeated starts, stops and reversals.
Minutes on the line, days in the lab
Time is where testing becomes a manufacturing problem. A full production test takes 15–30 minutes per joint. A design-validation programme for a new actuator model takes 3–10 days. The two long items, the thermal run and the holding test, don’t fit in an end-of-line slot at all, which is why they’re done on samples rather than on every unit.
Chart 1 · Duration of joint-module test stages, in minutes (typical range)
Source: Econotests, robot joint module test system (production test 15–30 min per joint; thermal run 60–120 min; holding test 30–60 min; design validation 3–10 days). Bars show the typical range, not a minimum.
Run the arithmetic for a humanoid line. Take 30 joints per robot and 20 minutes per joint, and one robot’s worth of actuators eats ten hours of bench time. A plant targeting a few thousand robots a year needs a room full of benches, not one. That’s one reason Chinese automation suppliers now sell complete joint-module assembly lines with the test stations built in.
Measuring ripple and efficiency without fooling yourself
Two numbers are easy to get wrong. Torque ripple first. It’s a fast signal riding on a slow one, and a sluggish sensor simply averages it away. HBK’s guidance for e-motor testing pairs torque transducers of up to 0.02% accuracy with bandwidth up to 6 kHz, and acquisition at 2 million samples per second at 18 bit. It also warns against heavy filtering: a strongly filtered signal carries a large phase delay and loses the amplitude and frequency information you were trying to measure.
Efficiency is the other trap. A joint warms up, its grease thins, its losses shift. Measure it cold and you’ll publish a number the robot never sees. Teragon Research’s proposed humanoid actuator protocol tackles this by warming each joint for 3–5 minutes first and only logging data while temperature rises less than 0.5 °C per second, with true-RMS DC-bus power meters sampling at 1 kHz or more and the mechanical and electrical channels on the same clock.
Benchmarking against a human
Cobot test specs answer one question: does this joint match its datasheet? Humanoid developers also want to know whether it matches a person. Teragon’s November 2025 preprint, “Human-Level Actuation for Humanoids”, proposes a scoring system built on six factors – range of motion, a human-equivalence envelope of simultaneous torque and power, torque-mode bandwidth, task-weighted efficiency, thermal sustainability and loaded speed. It normalises everything to a 75 kg, 1.75 m adult.
The targets it lists are a sobering read for actuator designers. Ankle push-off during walking takes 90–143 N·m and 190–260 W at 8–12 rad/s, at the same time. Plenty of actuators hit one of those numbers. Fewer hit all three at once, continuously, without overheating. The paper sets out a method rather than test results for commercial joints, and it’s a single-author preprint, so treat it as a proposal – but it’s the clearest attempt so far to define what “human-level” should mean on a bench.
The standards are still in draft
No published standard tells a humanoid joint maker how to test and report. That is changing, fastest in China. MIIT released a humanoid standard system in February 2026 with “limbs and components” as one of six pillars, and in August opened a framework for public consultation that aims for more than 100 key standards by 2028. On 13 August 2026, a working group of 29 organisations – among them the Hubei Humanoid Robot Innovation Center, the Hubei metrology institute, Unitree and Xiaomi – kicked off a series of seven national standards on humanoid test methods, starting with general provisions, perception, decision-making and motion control.
Internationally, ISO 25785-1 is at committee-draft stage. It covers safety for industrial mobile robots with actively controlled stability, which includes legged machines and humanoids, and it’ll shape what a joint has to do when things go wrong rather than how fast it spins. Suppliers aren’t waiting. Korea’s HIGEN RNM says it built ISO 13849-1 functional-safety requirements into its six humanoid actuator platforms (60–348 N·m) and quotes a back-driving torque under 1 N·m “under specified test conditions” – a phrase that is doing a lot of work until those conditions are standardised.
What to ask a joint supplier
- Which tests run on every unit, and which only on samples. Thermal and life data almost always come from samples.
- The conditions behind each efficiency and torque figure: temperature, speed, warm-up, and whether the drive is inside the measurement.
- The reference torque sensor’s accuracy and bandwidth. A slow sensor hides ripple.
- Per-unit test records with a serial number, not just a type datasheet – so a field failure can be traced back to its bench result.
Go deeper: The Humanoid Actuation Report
249 pages on motors, reducers, screws, bearings, encoders and drives, the vendors behind every major platform, and the cost curves that decide adoption to 2036.
The Humanoid Actuation Report All joints & actuators articles All humanoid.guide reportsKeep reading · Joints & actuators
Related articles
FAQ
How is a humanoid robot joint tested?
On a dynamometer bench: the joint drives an in-line torque sensor and a load motor that can brake it, drive it or hold it, while a power analyzer, temperature probes and a data logger record everything. Typical tests cover torque accuracy, ripple, stiffness, backlash, efficiency, thermal behaviour, holding torque and bandwidth.
How long does it take to test one joint actuator?
A production end-of-line test typically takes 15–30 minutes per joint. Thermal runs (60–120 minutes) and holding tests (30–60 minutes) are usually done on samples, and validating a new actuator model takes 3–10 days.
Is there a standard for testing humanoid actuators?
Not yet a published one. China started work on seven national standards for humanoid test methods in August 2026 and aims for 100+ key humanoid standards by 2028. ISO 25785-1, on safety for legged and other dynamically stable industrial mobile robots, is at committee-draft stage.
Why does test temperature matter for efficiency?
Losses in the motor winding, the reducer and its grease all change as the joint warms up. A joint measured cold can look more or less efficient than it is in use, so good protocols warm the joint up and only log once temperature has stabilised.
Sources
- Econotests – Robot Joint Module Test System: Torque, Stiffness and Efficiency for Cobot Actuators (eight test items, acceptance values, test durations)
- Lanmec Technology – 1000 N·m robot joint module test bench (performance, durability, brake, overload and bus-communication tests)
- Magtrol – Robotics motor testing (M-TEST 7 test routines, dynamometers, torque transducers)
- HBK – Measuring Torque Ripple Accurately (webinar, 2022) (sensor accuracy and bandwidth, sample rate, filtering phase delay)
- Teragon Research – Human-Level Actuation for Humanoids (arXiv preprint, 2025) (test protocol, instrumentation, human torque and power targets)
- eWeek – China Proposes 100+ Humanoid Robot Standards by 2028 (MIIT standard system and consultation)
- SMM – Work launched on national standards for humanoid robot test methods (August 2026 kickoff, seven standards, 29 organisations)
- ISO – ISO 25785-1 (committee draft) (safety for industrial mobile robots with actively controlled stability)
- HIGEN RNM targets humanoid robotics’ next bottleneck: the actuator (60–348 N·m platforms, back-driving torque, ISO 13849-1)