A gauging request lands on the robot cell you are already building: check this slot, that bore, this width, without adding a station. The first question is accuracy, and the honest answer is a number, not a promise. A wrist-carried camera gauges like a good digital caliper: plan on about ±0.05 mm. Put the camera on a fixed stand and let the same robot present parts, and the budget roughly halves to ±0.02 mm. The three error terms behind those numbers are robot pose repeatability, lens perspective error and thermal drift. This guide works each one into a budget you can copy, then gives the decision for each tolerance class.
Caliper class, in numbers
A shop-floor digital caliper resolves 0.01 mm and repeats to roughly ±0.02 mm in practiced hands. That is the class a robot-carried camera competes in. Vision metrology applies a 10:1 working rule — measurement uncertainty at or below one tenth of the tolerance — and the vision metrology accuracy and repeatability rule shows how to budget toward it. Combining the two gives the targets:
- Camera on the wrist, about ±0.04 mm uncertainty: covers tolerances of ±0.4 mm and looser.
- Fixed camera station, about ±0.02 mm: covers ±0.2 mm and looser.
- Tighter than that: a dedicated gauge, not more camera.
The three error terms that set the budget
The cell gives you three terms to add. None of them is the image sensor — by the time the pixel check below is done, the camera is the smallest error in the chain.
Robot pose repeatability
Datasheets quote pose repeatability per ISO 9283: the spread over many returns to one commanded pose, not a maximum error. Mid-size six-axis arms typically print 20-50 µm. The term enters a gauge only where the camera's pose relative to the part changes between calibration and measurement — in practice, whenever the camera rides the wrist.
Absolute accuracy does not enter: an uncalibrated arm can sit 0.5-1 mm away from commanded position because of zero and kinematic errors, yet repeat to datasheet, and the gauge only measures geometry inside its own calibrated image. Swap the architecture — camera fixed, robot presents the part — and this term leaves the budget; fixture placement spread replaces it, usually an order smaller.
Perspective error and calibration residual
A standard lens maps pose error into a height-dependent edge shift. The estimate: shift ≈ part height × tan(tilt). A 10 mm tall part at 0.2° of tilt moves edges by 10 mm × tan(0.2°) ≈ 0.035 mm; at 0.5° the shift is 0.087 mm and the whole wrist budget is gone. Telecentric optics, which accept only near-parallel rays, remove the term — see when you need a telecentric lens for gauging. The hand-eye calibration that ties image coordinates to the robot frame adds its own residual, typically 10-20 µm on a clean cell; robot-mounted camera calibration: capture and solve walks the capture and solve steps.
Thermal and mechanical drift
Steel grows about 12 µm per metre per °C. A 300 mm stand column therefore moves 3.6 µm per °C, so a 10 °C morning-to-afternoon swing is 0.036 mm — most of a fixed-station budget on its own. Hold the swing to 5 °C and the term drops to 0.018 mm. Focus moves with the same heat and drags the image scale along (camera focus drift with temperature: diagnose daytime blur), and a bumped fixture moves the calibration the same way (static camera calibration accuracy: why moved parts drift off target).
Worked error budget: slot width at ±0.5 mm
The part is a machined bracket with a ±0.5 mm slot width tolerance; the robot carries the camera over the fixture and dwells before each trigger. Inputs, in mm, from the three terms plus the pixel check below:
- u_r — robot pose repeatability: 0.03
- u_c — hand-eye calibration residual: 0.015
- u_t — thermal drift after a 5 °C soak: 0.018
- u_i — image processing: 0.002
The 10:1 target for a ±0.5 mm tolerance is 0.5 mm / 10 = 0.05 mm. The budget lands at ±0.038 mm: pass, with about 24% margin.
Pixel, focal length and depth checks
The example sensor reads 3072 pixels across a 60 mm field of view: 60 mm / 3072 px = 0.0195 mm/px, and a subpixel edge fit at 0.1 px contributes 0.002 mm — the u_i term. The camera resolution calculator runs this for any sensor and FOV pair.
Magnification sits near 0.1× (a 60 mm FOV on a sensor roughly 6 mm wide), so f = WD × m / (1 + m) = 300 mm × 0.1 / 1.1 ≈ 27 mm; a 25 mm lens lands slightly wide, which is fine with margin. Depth of field, estimated as DOF ≈ 2·N·c / m² at f/5.6 with c = 2 px × 0.02 mm = 0.04 mm object-side, gives 2 × 5.6 × 0.04 / 0.1² ≈ 45 mm — a ±1 mm placement spread sits deep inside it. Depth is not the limiting term; pose tilt is. The depth of field calculator repeats this for your aperture, and the focal length calculator below solves f from sensor and FOV.
Trigger discipline: dwell 200 ms after the in-position signal, then fire a 1 ms strobed exposure. Vibration left in the arm then cannot reach the edges.
Verdict for this case: the wrist-carried camera holds the tolerance. If the drawing said ±0.1 mm, the 10:1 target would be ±0.01 mm and the same budget misses by nearly 4× — the architecture, not the camera, has to change.
Fixed station, wrist or dedicated gauge
| Set-up | Realistic uncertainty | Covers at 10:1 | Choose when | Watch |
|---|---|---|---|---|
| Camera on the wrist | ±0.04 mm | ±0.4 mm and looser | many features, mixed parts, one camera | pose tilt, dwell before trigger |
| Camera fixed, robot presents | ±0.02 mm | ±0.2 mm and looser | one or two tight features, high takt | fixture spread, thermal drift |
| Dedicated air or drop gauge | 1-5 µm | ±0.02 mm and tighter | one bore or diameter drives scrap | plug wear, jet calibration |
The fixed-station row drops the robot from the measurement chain entirely: no repeatability term, no hand-eye residual — just scale calibration (0.01 mm), thermal drift (0.018 mm at a 5 °C swing) and image processing (0.002 mm), which combine as sqrt(0.01² + 0.018² + 0.002²) ≈ ±0.02 mm.
Air gauging reads through non-contact jets in the bore and resolves diameter to a few micrometres, an order finer than the camera rows. Keep the robot camera for outer features and handling, and let it ferry parts to the air plug when the bore is the scrap driver.
The set-up that holds the budget
For the measurement stop, pick a global-shutter camera with the smallest pixels your FOV allows: a 5 MP colour sensor in 1/1.45" format over Fast Ethernet covers the ±0.5 mm class, and an 8.8 MP, 1" C-mount model leaves room to tighten the FOV later. Both fit this cell:
- Interface
- Fast Ethernet
- Resolution
- 5 MP
- Interface
- USB 2.0
- Resolution
- 6-10 MP
When the cell needs pass/fail screening rather than numbers at one station, a fixed-focus vision sensor with a 240 mm focus distance does that at the lowest frame cost:
- Interface
- Fast Ethernet
- Resolution
- Up to 3 MP
- Sensor format
- 1/6"
- Shutter
- Global
The controller closes the loop: trigger, strobe, and the result to the PLC. PoE camera ports power the camera, and eight built-in light channels drive the backlight without separate dimmers:
- LAN ports
- 5+ LAN ports
- USB 3 ports
- 3-4 USB 3 ports
- PoE
- PoE camera ports
- Light outputs
- 8 built-in channels
For the fixed station, the bench stand holds the camera square over the fixture — 700 mm of column height on a 400 x 450 mm base:
- Type
- Camera stands
Pair the gauge camera with a telecentric lens for wrist work or a locked-focus fixed-focal lens on the stand, light edges with a strobed collimated backlight, and set f/5.6-f/8 for the depth of field you computed. Browse the smart cameras range or the camera finder with your pixel-size filter, and match the compute to the camera over in vision controllers.
Checks before you commit
- Dwell and trigger: 200 ms settle after in-position, 1 ms strobed exposure; no measurement while the arm still moves.
- Gauge R&R in the cell: 30 pick-and-measure cycles, with the 6σ spread at or below half the budget.
- Master part at every shift start; if it reads more than half the budget off, recalibrate that day, not that quarter.
- Log cell temperature; keep the swing within 5 °C or compensate against the master.
- After any collision, gripper rebuild or joint replacement, redo the hand-eye calibration — restore vision accuracy after a robot joint replacement covers the recovery path.
Decision, case by case
- Tolerance ±0.5 mm or looser on mixed parts: camera on the wrist, telecentric lens, ±0.04 mm budget. Proceed.
- Tolerance around ±0.2 mm on one or two features: fixed camera on a stand, robot presents parts, the fixed-station budget from the table above.
- Tolerance ±0.1 mm or a tight bore: dedicated gauge in the same cell; the camera keeps handling and the loose features.
More cell-planning material sits on the applications topic page. Size the camera and controller from the categories above, stand the fixed station on the camera stands range, or start at the system configurator. If the budget does not close for your drawing, contact us with the tolerance and the feature list.
Questions
Does robot absolute accuracy matter for camera gauging?
No. A gauge measures part geometry inside one calibrated image, so it rides on pose repeatability, not on where the arm sits in the world. An uncalibrated six-axis arm can miss commanded position by 0.5-1 mm from zero and kinematic errors yet still repeat to its datasheet 20-50 µm. Absolute accuracy only decides whether the part lands inside the field of view; repeatability sets the scale error.
What tolerance can robotic camera gauging still cover?
Apply the 10:1 rule: uncertainty no more than a tenth of the tolerance. A wrist-carried camera at ±0.04 mm therefore covers tolerances of ±0.4 mm and looser, and the fixed-station budget of ±0.02 mm covers ±0.2 mm. For ±0.1 mm or tighter, budget a dedicated gauge instead; stretching the vision budget moves the argument from the cell to the calibration lab.
How often should I recalibrate a robot-mounted gauge?
After any collision, gripper rebuild, joint replacement or camera remount, and otherwise on a temperature or seasonal schedule. Between events, run a master part at each shift start: if it moves more than half your budget, recalibrate that day, not that quarter. A hand-eye calibration on a clean cell typically lands between 10 and 20 µm of residual.
Can I gauge while the robot is moving?
Not for dimensional work. Motion blur and rolling-shutter skew put tens of pixels of edge error into an image that the whole budget fits inside two pixels. Stop the arm, dwell about 200 ms after in-position, then trigger a strobed 1 ms exposure. Presence and code checks tolerate motion; width, position and diameter gauging do not.
Why is an air gauge better for bores?
A bore is the worst case for a camera: the edge sits inside a deep, often shiny cylinder, so lighting and perspective errors grow exactly where the tolerance is tightest. Air gauging flows non-contact jets through the bore and reads diameter to a few micrometres, an order finer than the camera budget. Keep the robot camera for outer features and ferry parts to the air plug for the tight bore.




