Why a humanoid joint needs a health record

Joints are the most numerous wear parts on a humanoid. Each contains a motor, a reducer, bearings, grease, seals and one or two encoders, and each of those ages at a rate set by load, speed, temperature and how often the robot falls. A catalogue L10 life tells you how a population of reducers behaves on average. It does not tell you that the left knee of robot 14 is wearing faster than the rest of the fleet.

Condition monitoring closes that gap. The general method is set out in ISO 17359, the guideline for condition monitoring and diagnostics of machines: pick the failure modes that matter, choose measurements that change as those modes develop, set baselines and alarm limits, and act before the part fails. For a fleet operator the payoff is planned rather than unplanned downtime. Agility Robotics’ Arc fleet platform for Digit, launched in March 2024, already reports uptime and mean time between incidents (MTBI) per deployment. Joint-level prognostics is the next layer down.

What wears out first in a strain-wave joint

Monitoring starts with failure modes. The most detailed public analysis for the reducer used in many humanoid arms and some legs comes from Politecnico di Torino, which ran a failure mode, effects and criticality analysis (FMECA) on the harmonic drives of a Universal Robots UR5 cobot. Each mode was scored 1–10 for severity, occurrence and detectability; the product is the risk priority number (RPN). Wear at the interface between wave generator and flexspline scored highest at 180, followed by a crack at the flexspline’s rear cross-section (162) and lubricant degradation (160). Buckling and ratcheting, the dramatic overload failures, scored lowest because they are rare in a correctly sized joint.

The ranking matters for monitoring because the top three modes all develop slowly. Interface wear and grease breakdown raise friction and lost motion over weeks or months; a fatigue crack grows cycle by cycle. Slow faults are the ones that can be caught. Overload failures from a fall, covered in overload protection, happen in milliseconds and are a design problem, not a monitoring one.

Five signals a joint can report

• Motor current, read as torque. Every servo drive already measures phase current for control, and TI lists current sensing on all joints among the core sensing functions of a humanoid. FANUC’s ZDT diagnoses reducer deterioration on each axis from motor torque. The catch, noted by the Torino group, is that the control loop hides wear by simply pushing more current, so the signal has to be compared with a model or a baseline.
• Temperature. Winding and drive temperature protect the motor from the thermal wall. For wear, temperature is a slow, supporting signal: it moves late, and ambient conditions and duty cycle move it too. Its main job in monitoring is to correct friction, which changes strongly with grease temperature.
• Two encoders. A joint with encoders on both sides of the reducer can compare motor angle divided by ratio with output angle. Growing lost motion or changed torsional stiffness points to wear between wave generator and flexspline.
• Vibration. On a harmonic-drive test rig, Huddersfield researchers found that drives with flexspline tooth wear had lower RMS vibration but higher kurtosis than healthy drives, with abnormal peaks near the fault frequency (49.5 Hz in theory for their setup). Vibration needs an accelerometer in or near the joint, which most humanoid joints do not yet carry.
• Debris in the grease. Nabtesco, the leading maker of RV reducers for industrial robots, patented a particle detector for reducer casings: conductive magnets attract wear particles and double as electrodes, and when enough debris bridges the gap the resistance drops and an alarm fires. It is a direct measure of metal loss, at the cost of a sensor per joint.

Friction is the wear gauge

Of these signals, friction estimated from motor torque is the one that needs no extra hardware. The classic work comes from André Carvalho Bittencourt’s 2014 doctoral thesis at Linköping University, done with ABB Corporate Research. Wear changes friction in a joint, but so do temperature and load, often by more. His friction model accounts for all three, so the wear component can be separated even when load and temperature vary widely, and it was developed and tested in many thousands of experiments. Without such a model, a simpler method still works: run the same motion repeatedly and compare torque with data recorded when the robot was healthy. Those test runs can be scheduled in idle time, for example between shifts.

What the industrial robot makers already run

Fleet-scale joint monitoring is a solved product in factory robots, and the numbers show what humanoid fleets could expect. FANUC’s Zero Down Time service connects more than 40,000 robots, diagnoses reducer deterioration per axis from motor torque, detects reducer abnormalities several weeks before failure and, by FANUC’s count, had prevented over 4,000 downtime cases by July 2025. ABB’s condition-based maintenance service, launched in January 2021, compares duty, speed, acceleration and gearbox wear with ABB’s worldwide robot database and estimates the likelihood and timeframe of a failure for each robot.

Why humanoids make it harder

The industrial methods lean on one convenience: a welding robot repeats the same path thousands of times a day, so today’s torque can be compared with last month’s on an identical motion. A humanoid running a learned policy rarely moves the same way twice. Its legs see impact loads at every step, its load changes with every box, and its joints heat and cool with the task. That pushes humanoid monitoring towards the two approaches Linköping described: explicit friction models that correct for temperature and load, or scheduled self-test motions in idle time, such as a short routine on the charging dock.

Falls add a second category of event. A hard fall can damage a reducer in one shot, so a useful health record also logs peak torques and impacts per joint, not just slow trends. And because most humanoid joints carry no accelerometer or debris sensor today, current, temperature and encoder data will do most of the work. Dual-encoder joints, which some makers already fit for calibration and stiffness compensation, give monitoring a second, independent view of the reducer for little extra cost.

What to ask an actuator or robot supplier

• Which health signals does each joint export? Current, winding temperature, output encoder, peak-torque counters, fault logs, and at what rate.
• Is there a self-test motion? A repeatable routine, and a stored healthy baseline per joint from end-of-line testing.
• How is friction corrected for temperature? Grease friction changes with temperature; a wear alarm that ignores it will either miss wear or cry wolf on cold mornings.
• What lead time and false-alarm rate has been measured? On which failure modes, and on how many joint-hours.
• Can a single joint be swapped in the field? Monitoring pays off only if the flagged actuator can be replaced quickly.

FAQ

Sources