Robot wrist camera mounts: stiffness, flange fit and vibration

A camera on a robot wrist turns the bracket into part of the measuring chain: anything that flexes or rings comes back as pose error and smeared frames. The build question is mechanical, not optical — is printed plastic stiff enough for a wrist-mounted camera, or does the part need to be machined from aluminium? Below is the paper check that answers it with numbers you can run in ten minutes, the flange pattern and bolt torque to copy, and the material call for each case.

The decision, up front

Design the mount so its first resonant frequency lands at least four times above the arm's motion band. A collaborative arm drives its wrist with move-and-stop profiles whose content sits mostly below about 50 Hz, so the target is 200 Hz or better. A solid, 100%-infill PLA bracket that keeps a 200 g camera within 30 mm of the flange face reaches about 265 Hz in the worked example below — it passes, and printed plastic is the right answer. The same camera on a 2 mm-deep printed plate or hung 60 mm out lands inside the band and rings at every stop; shorten the lever, deepen the section, or machine metal. Once the camera is mounted, the tool-frame offsets that work on a gripper camera take over from the mechanics.

What the arm excites

A wrist joint spends its life accelerating a camera from stop to stop. The commanded joint paths put most of their energy below roughly 50 Hz on collaborative-class arms: modest speed limits, direction reversals, and the deceleration overshoot at each stop. A mount mode inside that band is driven at resonance on every cycle, and the camera orientation oscillates for tens of milliseconds after the arm stops — long enough to blur a triggered frame or bias a pose estimate. A mode four times higher barely responds, which is where the 200 Hz target comes from. Plastic also helps: PLA damps vibration far better than aluminium, so an excited plastic bracket stops ringing quickly, while a slender aluminium lever can hum after each stop.

The paper check: resonance and deflection

Model the bracket as a cantilever: the flange face is the root, the camera centre of mass hangs at lever length L, and the stiffness that matters is bending along the load direction. Three lines of arithmetic give you the answer.

Worked example — solid PLA lever, 30 mm long, section 20 mm wide by 10 mm deep, camera 200 g (m = 0.2 kg), PLA modulus E ≈ 3 GPa for a solid print:

• I = 0.020 × (0.010)³ / 12 = 1.67×10⁻⁹ m⁴
• k = 3 × 3.0×10⁹ × 1.67×10⁻⁹ / (0.030)³ = 15 / 2.7×10⁻⁵ ≈ 5.6×10⁵ N/m
• f₁ = √(5.6×10⁵ / 0.2) / 2π = √(2.78×10⁶) / 6.28 ≈ 1666 / 6.28 ≈ 265 Hz

The bracket's own mass is about 7.5 g and adds almost nothing. At a hard 2 g stop (a = 19.6 m/s²) the tip bends δ = 0.2 × 19.6 / 5.6×10⁵ ≈ 7 µm — micrometres against depth-camera pixels measured in fractions of a millimetre at close range. That is a passing mount.

Two ways to fail the same check:

• Shallow section: the same lever 2 mm deep gives I = 1.33×10⁻¹¹ m⁴, k ≈ 4.4×10³ N/m, f₁ ≈ 24 Hz — square in the motion band — and a 0.9 mm tip sway at 2 g. Section depth enters with the cube, which is why thin, shelled prints wobble and crack at the bosses where solid blocks do not.
• Long lever: stretch the lever to 60 mm and k falls with the cube of L, so f₁ drops by √8, from 265 Hz to about 94 Hz. Length hurts more than anything else.

What moves the answer: PLA's modulus spans roughly 2.5–3.5 GPa, which moves f₁ about ±9%. Doubling camera mass divides f₁ by √2 (0.71). On this lever the 200 Hz target holds up to m = k/(2π·200)² ≈ 0.35 kg; a 0.5 kg camera lands near 168 Hz and needs metal or a shorter lever.

The flange: pattern, screws and torque

The tool side of most six-axis arms follows ISO 9409-1, the standard family for robot mechanical interfaces. The common collaborative wrist size is the 50 mm version: four M6 tapped holes equally spaced on a 50 mm bolt circle, around a central pilot bore. Model that interface once, and every camera bracket starts from the same plate. Typical manuals for this interface specify M6 screws of strength class 8.8 tightened to 8 Nm — copy that value, and put a stiff steel washer under every head so the clamp load spreads instead of digging into plastic.

Two rules keep the joint honest: bolt through the bracket where you can, so the screws clamp washer-to-washer rather than threading into printed plastic; and if a plastic boss must take a screw, press in a brass insert rather than threading the raw print. For the camera's own foot, most depth cameras carry a 1/4-20 UNC tripod boss — same logic, insert or through-bolt. Camera makers publish STEP files of their housings, and the housing workflow that starts from maker STEP files plugs straight into this flange model.

Cables, because they push back

A stiff bracket still loses to a taut cable. Leave a service loop between the forearm and the camera, anchor it to the arm body, and make sure no cable tension ever reaches the bracket or the camera boss. On an Ethernet camera, the GigE camera network setup on a robot wrist covers the connector and strain-relief side in detail.

PLA, PETG or aluminium

Print PLA solid — 100% infill on a bracket this small costs minutes — with four or more perimeters, and orient the part so the lever's tension face runs along continuous walls rather than across layer lines. PLA softens around 55–65 °C, so keep it away from warm wrist housings and hot cells. PETG and glass-filled nylon buy heat margin and tougher layers at similar stiffness. Aluminium buys about 23 times the modulus of PLA, near-zero creep and no thermal softening, at the cost of machining and lower damping.

Creep is the quiet one: PLA under sustained bolt preload and cantilever load slowly relaxes, and the camera orientation drifts by fractions of a degree over weeks. That shows up as stale extrinsics — the IMU yaw offset after robot mounting and the robot-mounted camera calibration workflow both live downstream of exactly this drift.

Set-up checks before you run

1. Torque the four M6 screws to 8 Nm with washers, and re-check after the first day of cycling — plastic joints relax.
2. Flick test: rap the bracket with a knuckle and listen to how the sound dies out (the pass and fail signs are described in the FAQ below).
3. Stop-and-capture: trigger at every stop and add settle time until no frame smears — the same vibration-and-blur trade covered in autofocus versus fixed focus on moving robots.
4. Repeat-pose spread: approach a fixed target twenty times; a stable spread run to run is stiffness, a growing one is creep or a loose bolt.

What to mount, and where to develop it

Compact 3D sensors suit this kind of bracket:

These Hikrobot models cover the three common wrist jobs — a ToF sensing camera for coarse localisation, projected structured light for close-range surface work, and a laser-galvo stereo camera for scanning from the wrist. Keep the optical geometry in mind when you pick one: stereo camera spacing for robots trades depth quality against bracket width, and the 3D camera working distance guide sets how far off the flange the sensor actually has to sit.

Develop the cell on the bench before the arm: a camera stand holds the sensor still while you fix working distance, focus and lighting, so the wrist only ever replays a proven geometry.

When the sensor goes onto the arm, the wrist camera bin-picking limits show what one snapshot from that mount can and cannot decide.

When the usual answer is wrong

The printed-plastic answer flips in three cases. First, the view demands the camera hang 60–100 mm off the flange for clearance or field of view: frequency falls with L^1.5, so add an aluminium strut or a triangular gusset back toward the flange root rather than hoping. Second, the wrist runs warm — PLA at a 55 °C softening floor is the wrong polymer, move to filled nylon or metal. Third, the process measures micrometres through the camera pose, not millimetres: any creep shows, so metal it is. Outside those three, print plastic, keep it short, keep it solid.

Next step

Pull the flange dimensions from the arm manual, run the three-line check above, and build to that number. The 3D sensors finder filters wrist-capable cameras by type, and the stand configurator specs the bench stand you develop on. More on this in the 3D imaging topic or the guides hub.