Most "the joint doesn't meet spec" conversations we have end at the mounting interface, not inside the joint. That's good news — installation problems are cheap to fix and cheaper to prevent — but only if you know where to look. Here is where arcsecond-class performance actually leaks out of a machine, in descending order of frequency.
The flange is part of the gearbox
A strain wave gear's flexspline is a precision spring a fraction of a millimetre thick. Bolt its housing to a surface that isn't flat and you preload that spring unevenly: the housing becomes slightly elliptical, and with it the tooth engagement you paid for. The symptoms are a once-per-output-revolution ripple in torque or noise, and backlash that seems to vary with position.
- Machine the mating face; don't mount precision joints to as-cast or painted surfaces.
- Respect the flatness and screw specifications in your frame's installation drawing, and if a spec is missing from your paperwork, ask us rather than improvising.
- Check for trapped debris and raised burrs around tapped holes before mating. A single chip under a flange is a machined-in error.
Bolt discipline
Fastener errors mimic gearbox faults convincingly:
- Torque in a crossing pattern, in stages. Running one bolt fully home first tilts the housing and locks the distortion in.
- Use the specified grade and length. Too long bottoms out and lifts the flange; too short strips under the rated preload.
- Thread-lock, not hope, on anything that reverses load: a joint that reverses thousands of times a day is a bolt-loosening machine by design.
The output side: stiffness you can throw away
The joint's torsional stiffness only matters if the bracket you bolt to its output is stiffer. A sheet-metal arm on a precision joint moves the compliance problem downstream where no gearbox can fix it; the stiffness budget is a series chain, and the softest element owns the result. Design output structures at least as stiff as the joint, and mount payloads with located fits (a pilot bore or dowels), not friction and optimism.
Cables: the forgotten load
An integrated joint's cable carries power and a fieldbus, and on a moving axis it flexes with every cycle. Unmanaged, it becomes three faults at once: a parasitic torque the current loop can't explain, a fatigue failure scheduled for mid-production, and — on force-sensing axes — a phantom force that drifts with pose. Route cables along the neutral axis of rotation, clamp at both ends with service loops, and respect the cable's own bend radius as a spec, because it is one.
The first hour: commissioning checks that pay forever
Before the machine does its first useful work, four measurements turn "it seems fine" into a baseline you can defend:
- No-load current sweep. Rotate each axis slowly through full travel, unloaded, and log the current trace. It should be flat and symmetric; a position-dependent ripple is a mounting distortion talking. This is also your wear baseline for every service interval after.
- Backlash spot-check. Compare against the as-shipped value logged to the unit's serial number. Installed backlash should match the bench. If it doesn't, the difference came from your interface, and now is the hour to find it.
- Thermal soak. Run the expected duty cycle for long enough to reach steady state and record the housing temperature. This validates the thermal budget against reality, not the spreadsheet.
- Repeatability run. Ten approaches to the same target from both directions, measured at the tool. This is the number your process will actually live with; file it next to the serial number.
The short version
Machined flat mounting faces, staged crossing-pattern bolt torque, output structures stiffer than the joint, cables treated as components, and a first-hour baseline of current, backlash, temperature, and repeatability. None of it is exotic; all of it is the difference between the spec you bought and the spec you get.
Related: How harmonic drives fail · Reading the datasheet.