How long does a humanoid joint last? The lifetime problem nobody prints on the spec sheet

Rated life is a bearing number, not a robot number

When a reducer maker quotes a lifetime, it is quoting a statistical bearing rating. L10 is the running time by which 10 percent of units are expected to have failed under stated conditions; L50 is the point where half have. For its CSD and SHD strain-wave gears, Harmonic Drive publishes an L10 of 7,000 hours and an L50 of 35,000 hours, both defined at rated torque with the input turning at 2,000 rpm. Those two numbers frame the honest range: a hard-driven joint is a wear part on a horizon of thousands of hours, and a gently driven one can plausibly run for a robot’s whole service life.

Which of those two joints you get is set by a cube law. The catalog life formula scales expected hours with the cube of the ratio between rated and actual average torque (and linearly with speed). Run the same gear at half its rated torque and expected life multiplies by eight; run it 25 percent over rating and life roughly halves. This is why actuator sizing, not component quality, is the first reliability decision a humanoid designer makes — and why robots that routinely operate near peak torque, as legged machines do when they squat, climb and catch falls, consume gear life far faster than an arm bolted to a table.

How a joint actually dies

A 2024 overview of harmonic-drive failures in industrial robots, published with the PHM Society, maps the wear sites precisely. The gear’s end of life is defined as the onset of pitting in the inner race of the wave generator — the elliptical bearing at the heart of the drive. The most common catastrophic mode is fatigue fracture of the flexspline, the thin flexing cup, with cracks nucleating at the tooth roots. Wear between the wave generator’s outer race and the inside of the flexspline quietly erodes efficiency long before anything breaks, and tooth wear concentrates near the flexspline’s input end because of the coning angle the cup adopts as it flexes. In the study’s risk ranking, that interface wear scored highest of all failure modes.

Two slower processes matter just as much for a walking robot. First, backlash grows: a 2023 study in Engineering Failure Analysis modelled how harmonic-gear backlash degrades with accumulated tooth wear and assembly error, which is felt as a robot that gradually loses positioning crispness. Second, shock overload can make a strain-wave gear ratchet — teeth jump the mesh under a torque spike, damaging them permanently. Harmonic Drive publishes the threshold per size: 88 N·m for a size-14 CSD, the class of gear a humanoid might carry in a wrist or elbow. A bad fall can deliver that. Designers respond with backup ratio — the same overview recommends flexsplines with a rim backup ratio of at least 1.3 so a cracked rim fails safely; the UR5 gear examined measured 2.27.

Spec-sheet MTBF vs the factory floor

Robot makers advertise mean time between failures of 40,000, 60,000, 80,000, even 100,000 hours. The Robot Report’s review of field research covering more than 400 factories tells a different story: average robot-cell reliability of 88 percent and a mean time to failure of 87 minutes for the cell as a whole. The reconciliation is that 80 percent of stoppages come from everything around the robot — conveyors, sensors, tooling — and only about 2.5 percent of downtime traces to true robot mechanical failure in gears, reducers and bearings. Both claims are true at once: the transmission hardware is genuinely reliable, and the system built around it still stops constantly.

Humanoids inherit that gap and widen it. An industrial arm is bolted down, runs a fixed cycle inside its rated envelope, and never gets kicked. A humanoid carries its transmissions through impacts with the floor at every step, spends its torque budget near the limit whenever it squats or lifts, and — per the cube law above — pays for that duty cycle in gear life. The field record is also simply young: Digit’s June 2024 deployment at GXO near Atlanta was the first commercial humanoid deployment anywhere, so nobody yet has the decade of fleet data behind industrial MTBF claims. Teardown evidence points the same direction: Munro Live’s disassembly of the Unitree G1 found planetary joint actuators and heat-pipe cooling — engineering aimed squarely at the load and heat cycles that set joint life.

What builders do about it

The reliability playbook emerging across the industry is less about exotic materials than about arithmetic and architecture.

• Derate the joint. The cube law is the cheapest life-extension technology that exists: size the gear so that walking, the dominant duty, sits near half of rated torque, and reserve the rating for lifts and stairs. Eight times the life costs one size step in mass and money.
• Choose robustness over precision where you can. Quasi-direct-drive robots run low-ratio planetary gears with thick, simple teeth — one reason the approach keeps winning in legs, where impacts are constant. Strain-wave gears keep the wrists and arms, where loads are milder and precision matters more.
• Spread the load in linear joints. Where builders pay for planetary roller screws over ball screws, they are buying contact area: supplier comparisons credit roller screws with up to 15 times the dynamic load capacity and up to 15 times the travel life of a same-size ball screw, at 3–5 times the price.
• Validate before scaling. Schaeffler says its formed strain-wave gearboxes for humanoid joints completed validation ahead of scaled production planned for 2027 — a Tier-1 supplier treating humanoid gear life as a qualification problem, the way automotive components are qualified.
• Or make the joint disposable. Berkeley Humanoid Lite takes the opposite bet: 3D-printed cycloidal actuators in a robot costing under $5,000, where a worn gearbox is a print job, not a warranty claim.

FAQ

Sources