A robot joint backlash spec is usually one number on a datasheet standing in for three different measurements — backlash, repeatability and torsional stiffness — that fail in three different ways. Buyers compare that number across suppliers, suppliers quote it under conditions nobody states, and the joint that looked best on paper turns out to be the one the controller cannot hold still.
This is how we read those sheets before quoting, and it sits next to the torque side of the same decision in how to size joint modules. Torque gets a joint into the building; accuracy decides whether the arm is useful once it is there.
The three numbers behind one figure
Backlash, repeatability and stiffness answer different questions, and a sheet that gives you only the first is telling you the least interesting third of the story.
- Backlash (lost motion) — how far the output can be rotated by hand, with the input locked, before anything engages. A static measurement, made new, unloaded, at room temperature.
- Repeatability — whether the joint returns to the same output position after many moves, under load, hot. This is the number that decides whether a pick-and-place cell holds tolerance at hour six.
- Stiffness — how far the output deflects per unit of applied torque. Not an error at all, but a spring between your commanded position and the real one, and the reason a perfectly calibrated arm sags under its own payload.
| Number | What it measures | Typical published form | What the sheet leaves out |
|---|---|---|---|
| Backlash | Lost motion at the output with the input locked | arcmin or arcsec, often a single maximum | Load, temperature, and whether it is measured new or after run-in |
| Repeatability | Spread of return positions over many cycles | ± value at the output, sometimes at the motor | Unidirectional or bidirectional — the second includes the backlash |
| Stiffness | Deflection per unit torque at the output | Nm/rad or Nm/arcmin, if it is quoted at all | The torque it was measured at; harmonic stiffness is not constant |
What backlash looks like by transmission
The figure is largely set by the reduction principle, which is why it is one of the few specs you can sanity-check from the transmission type alone. Ranges below are what the suppliers we work with publish — approximate, and not a substitute for the datasheet of the specific frame you are buying.
| Transmission | Backlash, typical published range | Why it lands there | Axis where it fits |
|---|---|---|---|
| Harmonic (strain wave), precision grade | ≤1 arcmin (≈60 arcsec) | Near-zero by elastic preload of the flex spline, not by tight machining fits | Wrist, elbow, shoulder pitch on arms |
| Harmonic, standard grade | ≤3 arcmin | Same mechanism, looser preload, better efficiency and life | General arm axes, torso |
| Planetary | 3–15 arcmin | Gear mesh clearance; cost and shock tolerance come from the same clearance | Hip and knee push, cost-sensitive axes |
| RV (cycloidal) | ≤1 arcmin | Many contact points sharing load, so clearance can be held tight without preload | Heavy-load hip in large humanoids |
One harmonic line we carry publishes ≤60 arcsec on the output; planetary lines sit an order of magnitude looser. Treat every figure as approximate until you have the datasheet for the exact frame and ratio.
Why a good backlash figure still misses the target
Two mechanisms do the damage, and neither shows up in the number on the sheet.
Encoder placement. If the encoder is on the motor shaft — the usual arrangement on compact integrated joints — the gear train sits between the sensor and the load. The controller then believes a position the output has not reached yet, and every direction reversal has to consume the backlash before anything moves. An output-side encoder, or a dual-encoder arrangement, removes most of that error at the cost of wiring, cost and calibration. Ask which one you are being quoted; the difference is bigger than the backlash spec itself.
Compliance under load. A joint can be tight and still soft. Harmonic joints are compliant by design — that is what makes them shock-tolerant and backdrivable — and their stiffness falls as torque rises, so the deflection at 80% of rated torque is not the deflection implied by a no-load figure. On a 20 Nm joint the difference can be visible on a dial indicator while the backlash still measures as specified.
Stiffness is a control parameter, not a quality score
Stiffer is not automatically better. Stiffness sets how much of a disturbance the controller has to reject in software, and how much of an impact reaches the reducer.
- Low stiffness — the joint absorbs contact and the loop is easier to stabilise, but position under load depends on the payload, and force control has to model the spring.
- High stiffness — cleaner position tracking and easier trajectory planning, but impacts go straight into the gear teeth or the flex spline.
The flex spline is where that bill arrives. A single impact at roughly three times rated torque can consume an estimated 5–15% of a harmonic reducer's flexural fatigue life; ten such events and a large part of the service life is gone. That is why the same joint that is ideal in an arm is a poor choice in an ankle, and why we ask what the axis actually contacts before recommending a frame.
Which number matters on which axis
Most teams over-specify backlash on every joint and under-specify stiffness on the two that need it. A rough allocation, on the kind of humanoid most of our customers are building:
| Axis group | Spec that decides success | Spec that is usually oversold |
|---|---|---|
| Hip pitch / knee | Stiffness and shock tolerance; backlash secondary | 1 arcmin backlash quoted as a headline |
| Hip yaw / waist | Stiffness under asymmetric load, plus brake holding | Peak torque |
| Shoulder / elbow | Backlash and repeatability, both measurable at the output | Maximum speed |
| Wrist | Backlash, weight, and encoder placement | Payload rating without a lever arm |
| Hand and finger joints | Backdrivability and size; arcmin figures are noise at this scale | Any backlash number at all |
Where two frames sit close on torque, the accuracy numbers are usually what separates them — the cross-brand comparison by torque band puts the harmonic lines side by side on one scale, and the harmonic range carries the per-frame figures.
Five things to get in writing
None of these are adversarial questions. A supplier who has measured their joint can answer all five in one email; the ones who cannot have been copying a catalogue.
- Encoder location — motor side, output side, or dual. This changes the effective accuracy more than the backlash figure does.
- Repeatability direction — unidirectional or bidirectional, and at what load and temperature.
- Test conditions — new or after run-in, and whether the backlash figure is a maximum or a typical value.
- Stiffness curve — torsional stiffness at the output across the torque range, not a single no-load number.
- Who measured it — factory acceptance sheet with the serial number on it beats a brochure every time.
Rules of thumb
- Arms and wrists: buy the backlash figure, but ask about the encoder first.
- Legs: buy stiffness and shock tolerance; a 3 arcmin joint with a robust reducer outlasts a 1 arcmin joint with a thin flex spline.
- Anything that takes impact: derate for fatigue events, not just for torque.
- Comparing two quotes: normalise both to output-side numbers at rated load before you compare them at all.
- Prototype first: a dial indicator on the output flange, loaded and warm, settles in an afternoon what a datasheet argument will not settle in a month.
Send us the axis and the load case
Tell us the axis, the payload at the flange and what the joint will contact, and we will shortlist frames across the factories on our brand list at the accuracy grade that axis actually needs — with the encoder arrangement and acceptance sheet confirmed before anything ships.
Ask us to check a joint spec sheet
FAQ
Is lower backlash always better on a robot joint?
No. Backlash is bought with preload, and preload costs efficiency, heat and shock tolerance. A joint quoted at 1 arcmin is usually the right choice on a wrist or an arm axis, and the wrong one on a leg that takes ground impact. Match the figure to the axis, not to the catalogue ranking.
Why does my joint hold position but miss the target after a direction change?
That is lost motion, not drift. If the encoder sits on the motor side, the gear train fits between the sensor and the output, so every reversal has to take up the backlash before the output moves. Check where the encoder is mounted before blaming the controller.
What stiffness figure should I ask a supplier for?
Torsional stiffness at the output flange, in Nm/rad, quoted at your operating torque rather than at no load. Harmonic joints are deliberately compliant and their stiffness falls as torque rises, so a single no-load number tells you very little about the joint under a real trajectory.
Related product pages
Specs quoted from manufacturer datasheets; final figures confirmed at quotation.