A robot joint housing is a stiffness problem wearing a strength problem's clothes. Deflection in the load path becomes position error at the tool centre point, so design against compliance, not against fracture: closed box sections, ribs at 60 to 80% of the adjoining wall placed along the real bending path, and 2.5 to 3.5 mm walls. Cast the structure and machine the interfaces — reducer bores and their reference faces need ±0.02 mm with concentricity and perpendicularity inside 0.02 to 0.03 mm, which only comes from cutting bore and face in one setup off cast datum pads. A380 is the default because robot castings are machined heavily; ADC12 goes on thin covers and gripper bodies. At 1,500 to 20,000 units a year a single-cavity die at $3,000 to $12,000 usually beats billet machining on a ribbed housing, and almost never beats it on a flat plate.
A robot joint housing is one of the few castings where strength is almost never the binding constraint. The part will not break. It will deflect, and because it sits inside a kinematic chain, every micron of deflection is reported to the customer as position error at the tool centre point.
That changes how the part should be designed, and it changes which features are worth paying for. This guide covers the structural castings inside an articulated arm — joint housings, harmonic drive and cycloidal reducer cases, arm links, motor flanges and gripper bodies — and the decisions that actually move performance and cost.
The problem: stiffness per gram, inside a tolerance chain
Two requirements pull against each other on every moving casting in a robot.
Mass at the outer axes is paid for twice. A kilogram added at the wrist reduces rated payload directly, and it forces larger motors, larger drives and heavier structure in every joint below it. The compounding is real enough that teams routinely trade cost for grams at axes 4 through 6 and not at the base.
Stiffness cannot be traded away to get those grams back. A controller can compensate for a repeatable offset, but it cannot compensate for a structure that flexes differently depending on load and pose. Compliance in a joint housing shows up directly at the tool, and a structure with a low first natural frequency also forces the servo loop to be detuned, which slows cycle time for reasons the mechanical drawing never mentions.
So the design target is section stiffness per unit mass, not tensile strength. All four of our production alloys sit between 296 and 324 MPa tensile, and all of them have the same 71 GPa modulus. Alloy choice does not change stiffness at all. Geometry does, and casting is the process that makes good geometry cheap.
Where the accuracy actually lives
Repeatability is built from a chain of interfaces: reducer to housing, housing to link, link to the next joint. Each one contributes to the stack. The casting’s job is to carry those interfaces rigidly and present them to a machine tool in a repeatable way — not to be accurate itself.
As-cast accuracy is around ±0.1 mm for the first 25 mm of dimension, ±0.15 mm across the parting line and ±0.2 mm across a moving slide. That is fine for a rib, a wall or an envelope. It is useless for a bearing fit. The strategy is therefore always the same: cast the structure, machine the chain.
| Interface | Requirement | How it is held |
|---|---|---|
| Reducer or bearing bore | ±0.02 mm, H7 fit | Finish bored after casting, located from cast datum pads |
| Bore to mounting face | Perpendicularity 0.02–0.03 mm | Bore and face cut in one setup, 4 or 5-axis |
| Coaxial ends of an arm link | Concentricity 0.03 mm | Both ends machined in one fixture, no re-datuming |
| Mounting plane flatness | 0.02 mm per 100 mm | Face milled; as-cast flatness is 0.1 mm per 100 mm |
| Bolt pattern position | ±0.05 mm | Drilled and reamed from the machined datum |
| Motor pilot diameter | H7 fit | Machined; cored holes are pilots only |
| Envelope, ribs, cosmetic walls | ±0.1 mm | As-cast, no machining |
The single most important line in that table is the second one. A bore and the face it must be square to should be cut without releasing the part. When they are, the relationship is held by the machine’s own geometry. When they are not, it is held by the sum of two fixturing errors, and that is where concentricity callouts quietly fail at robot calibration.
This is also why cast datum features matter more on robotics work than on most castings. Clamping on as-cast draft surfaces means every part presents a slightly different reference. We cast defined datum pads and locating bosses so the fixture sees the same geometry on part one and part ten thousand.
Reducer housings specifically
Harmonic drive and cycloidal reducers are unforgiving of housing error in a way that ordinary bearing mounts are not, because misalignment does not just add friction — it changes the load distribution across the gear teeth and shortens life.
Three requirements come up on nearly every reducer housing we quote:
Bore roundness, not just diameter. A bore that measures correct on two axes and is 0.04 mm oval will pinch the outer race and raise running torque. Roundness should be called out separately where it matters; we hold 0.02 to 0.03 mm routinely and can go tighter on a dedicated fixture.
A face square to the bore, with enough land. The mounting face for the circular spline or the reducer flange needs perpendicularity to the bore axis and a continuous bearing land, not an interrupted ring crossing a cored pocket. Cored features on the opposite side frequently creep into the land area during weight reduction, and it is worth checking that in CAD before the die is cut.
Local stiffness around the bore. The housing wall immediately surrounding the reducer should not breathe under load. A thin wall with a bored hole in it deforms ovally; the fix is a thickened boss blended into the surrounding wall with ribs running out to the nearest structural section. Keep the boss blended rather than abrupt — an isolated thick ring freezes last and becomes the porosity that the bore then cuts into.
That last point is the recurring tension in reducer housings: the bore wants material around it, and material around it wants to be porous. The resolution is a boss that is thick where the bearing needs support and tapers into the wall, combined with gating that places the final freezing somewhere other than under the bore.
Rib design that earns its mass
Ribs are the main instrument for stiffness per gram, and most of the mass wasted on robot castings is wasted in ribs that are not on a load path.
The rules we work to:
- Rib thickness at 60 to 80% of the adjoining wall. Thicker than the wall and the rib root becomes the last place to freeze, which produces a sink mark on the opposite face and shrinkage porosity inside the junction.
- Ribs along the bending path, not distributed evenly. A rib grid that looks symmetrical in CAD is usually carrying mass that does nothing. We ask for the load case — payload, reach, worst-case moment, dynamic factor — during DFM so ribs go where the section is actually in bending.
- Closed sections beat open ones for torsion by an order of magnitude. An open C-section arm link is dramatically less stiff in torsion than a closed box of the same mass. Casting makes the closed box affordable; a fabrication usually cannot.
- Radius every rib root. With 2.5 to 3.5% elongation, cast aluminum tolerates stress concentration badly. A sharp rib root under vibration is a crack initiation site, and it is free to avoid.
- Draft on every rib face, typically 1.5 to 3°, which also means a rib is thinner at its tip than its root and that taper is structurally useful.
A 3 mm wall with a well-placed rib grid typically beats a 6 mm plain wall in bending stiffness at around half the mass, and it cools far more evenly, which keeps the part dimensionally stable.
What to cast and what to machine
The cost mistake we see most often is a blanket machining allowance across a whole casting. It adds cycle time, it adds distortion risk because material is removed unevenly, and it adds no accuracy anywhere it was not needed.
| Feature | Decision | Reason |
|---|---|---|
| Reducer and bearing bores | Machine | ±0.02 mm, roundness, concentricity |
| Reducer mounting face | Machine | Perpendicular to bore, in the same setup |
| Motor pilot and flange face | Machine | H7 fit and squareness |
| Structural bolt patterns | Machine | ±0.05 mm position from a machined datum |
| Sealing lands for O-rings | Machine | Ra 1.6 µm, flat within 0.05 mm |
| Cable pass-throughs | Cast | Radiused as-cast edge is better than a drilled hole needing a bushing |
| Cooling fins | Cast | No accuracy requirement, impossible to machine economically |
| Ribs, walls, envelope | Cast | ±0.1 mm is ample |
| Non-structural clearance holes | Cast | Cored to ±0.1 mm, no machining needed |
| Cast datum pads | Cast, then machine the ones used as references | They exist to make machining repeatable |
We leave 0.5 to 1.0 mm of stock on faces that will be cut and cast everything else to net shape. Marking the machined areas on the drawing before tooling design is the single most useful thing a robotics customer can do, because it drives the gating strategy as much as the machining plan.
Alloy, briefly
Since modulus is identical across the alloys, alloy choice on a robot casting is about machinability, wall thickness and heat — never stiffness.
| Part | Alloy | Why |
|---|---|---|
| Joint and reducer housings | A380 | Best machinability, clean threads, stable bores through a lot |
| Arm links and base castings | A380 | Machined interfaces at both ends, highest tensile of the four |
| Covers, shrouds, cable guards | ADC12 | Fills 1.5 to 2 mm walls other alloys freeze short of |
| Gripper bodies, encoder housings | ADC12 | Thin wall and low inertia matter more than machining behaviour |
| Joints that are also the motor heat path | A360 | 113 W/m·K against A380’s 96, sometimes removes a separate heat sink |
The alloy comparison covers the full property set. For most robot structures the answer is A380 and the conversation is short.
Does the volume justify a die?
Robot programs sit awkwardly for casting. A successful model might ship a few thousand units a year, not the hundreds of thousands that automotive tooling assumes, and customers often arrive expecting to be told they are too small.
Usually they are not. The arithmetic that matters is tooling amortisation against the number of machining setups avoided:
| Annual volume | Usual answer | Notes |
|---|---|---|
| 1–50 | Machine from billet | No tooling, fastest path to a working prototype |
| 500–2,000 | Case by case | Simplified single-cavity die plus more machining often wins on a ribbed housing |
| 2,000–20,000 | Production single-cavity die | The normal range for a robot model |
| 20,000+ | Multi-cavity | Typical for grippers, sensor housings and AMR components |
At 1,500 units a year, a $9,000 single-cavity die adds about $6 per part in year one and nothing thereafter. A complex ribbed housing that needs four setups from billet recovers that immediately. A flat mounting plate does not, and we will say so rather than sell the tooling.
One tooling decision is specific to robotics and worth raising early: build the likely revision zones as replaceable inserts. Bolt patterns, connector cutouts and cable entries are what change between robot generations. Flagged at design stage, a revision then costs an insert and a week instead of a new die and six.
Converting a machined prototype into a casting
Most robotics programs should validate in billet first and convert once the geometry settles. The conversion is a redesign, and these are the changes it always involves:
- Add 1.5 to 3° draft to every face parallel to the pull direction
- Replace sharp internal corners with radii, especially at rib roots and boss blends
- Convert uniform thick walls into 2.5 to 3.5 mm walls plus ribs
- Hollow out solid volumes with cored pockets, opened in the pull direction where possible
- Thicken locally only where bearings, gearboxes and fasteners load the structure
- Move undercuts out of the geometry, or accept a slide and its $1,500 to $4,000
- Add cast datum pads sized for the machining fixture
- Keep the validated interfaces and envelope exactly as tested
Done properly the cast version is lighter than the machined original it replaces, because it is finally allowed to have the geometry the load case wanted all along.
If you want this reviewed against a real part, send the 3D model, the load case and the annual volume. You will get piece pricing, tooling cost, a machining plan for the bearing and mounting interfaces, and a written DFM report on where mass can come out without losing stiffness.
