Humanoid robot development programmes tend to over-specify machined parts in the wrong places. A finger link comes back with ±0.005 mm on every dimension and still will not assemble, because the error that mattered was flatness on a mounting face, and nobody specified it.
This post is about allocating tolerance rather than tightening it.
A robot is a kinematic chain, so errors add
Every joint in a chain contributes position error at the tool centre point. If six joints each contribute 0.02 mm of play, the end effector is not 0.02 mm off, it is up to 0.12 mm off in the worst case — and worse, the error changes with pose.
That changes what a machined part should be judged on. On a single part, a 0.01 mm dimensional error is trivial. On the twentieth part in a chain, it is not.
Practical rule: tighten the parts closest to the end effector, and the interfaces that repeat. A wrist flange that locates a gripper deserves more attention than a structural bracket in the middle of the arm.
Flatness and position fail differently
These two are often treated as interchangeable. They are not:
- Position error moves a feature. A bolt hole 0.05 mm off still accepts the bolt; the joint just sits slightly shifted.
- Flatness error tilts a joint. A mounting face with 0.02 mm of bow over 80 mm creates an angular error of roughly 0.014° — small, but it repeats at every interface and it loads one side of the joint.
In robot assemblies, flatness errors are the ones that produce visible wobble, uneven bearing preload and accelerated wear. They are also the ones most often missing from a drawing.
Thin walls: clamping is the error source
Anthropomorphic robot parts are weight-driven, which means thin walls. Thin walls deflect under clamping force, get machined while deflected, and spring back on release. The result is a part that measures correctly while held and incorrectly when free.
What actually works:
- Light clamping with a support medium. Low-melt fixturing or a potted fixture removes the deflection instead of fighting it.
- Symmetrical material removal. Rough both sides, stress-relieve, then finish both sides. Finishing one face completely before touching the other guarantees a bow.
- Measure free-state. Any dimension taken while the part is clamped is an in-process number, not a part number.
- 7075-T6 versus 6061-T6. 7075 machines cleaner and holds a better finish, but it is less forgiving of residual stress and more prone to distortion after heavy pocketing. For thin, heavily pocketed parts, 6061 sometimes holds form better despite lower strength.
Interface fits: where the assembly actually lives
The joint between two machined parts is where tolerance stops being abstract. Three interface types cover most robot joints:
| Interface | Typical fit | What decides success |
|---|---|---|
| Bearing seat | H7 / h6 or a light press | Roundness and cylindricity, not diameter |
| Dowel location | H7 / m6 | Hole position relative to the functional datum |
| Bolted face | Clearance holes | Flatness and bolt-pattern position |
A bearing seat held to nominal diameter but with 0.008 mm of ovality will still produce a loose or seized bearing — the diameter is right, the form is wrong. This is why roundness belongs on the drawing even when diameter is already specified.
Datum schemes that survive manufacturing
A datum scheme that requires the part to be perfectly fixtured to be inspected will not be used. The scheme that works:
- Pick one primary datum: the face that mates with the next part in the assembly.
- Make the secondary datum a feature that is machined in the same setup as the primary.
- Express everything else as position or profile relative to those two.
If the drawing needs a third datum to describe a functional relationship, that is a signal the design has an unnecessary intermediate interface.
Surface finish at the places that slide
Robot joints slide and rotate. Two surfaces matter:
- Bearing and bushing seats — roughness drives fit and, in a press fit, the actual interference. Ra 0.4 µm is the usual target; Ra 1.6 µm will change the fit.
- Sealing and wiper surfaces — a rotary seal on a turned surface with a helical tool mark will leak along the helix, not around the circumference. The fix is a plunge or a low-feed finish pass, not a tighter tolerance.
Inspection that reflects the requirement
A CMM report listing 40 dimensions is not an inspection plan. The plan that catches robot-part failures:
- Position of each interface feature relative to the primary datum, on a CMM with a stated alignment.
- Flatness of every mounting face, free-state.
- Roundness / cylindricity on bearing seats.
- Wall thickness measured at several points on thin shells, not one.
- First article approved before the batch is released, with values, not a pass stamp.
The short version
- Allocate tolerance by position in the kinematic chain, not uniformly.
- Flatness fails assemblies more often than dimensional tolerance.
- Thin walls: control clamping, remove material symmetrically, measure free-state.
- Bearing seats live or die on roundness, not diameter.
- One primary datum, one secondary machined in the same setup, everything else relative.
The process detail behind this is on our humanoid robot parts manufacturer page. We machine finger joints, end-effector mounts and sensor housings in 7075-T6, 6061-T6, TC4 and PEEK on 100 five-axis machining centres, to ±0.005 mm, with first article approved before any batch ships.
If you are allocating tolerance across a hand or wrist assembly and want a second opinion on where it should go, describe the chain in the comments.
Top comments (0)