Vision guidance positioning error budget for ±0.5 mm targets

A drawing that calls out ±0.5 mm for a vision-guided weld, bead or pick is not asking whether a camera is accurate in general. It asks whether four separable error sources — calibration, model, locating and picking — sum to less than the tolerance in your cell. This guide splits the budget into those four stages, gives each one a numeric target, and works a spot-weld example end to end with the arithmetic shown. You finish with a target per stage and a test that confirms each before the robot touches a part.

The four stages of a guidance error budget

Every vision-guided move inherits error from four places. Calibration error is the gap between what the camera reports and where the part truly sits in the robot frame. Model error is the gap between the taught model and the part in front of the lens, including the part's own tolerance. Locating error is how precisely the image measurement finds the feature. Picking error is the arm and gripper repeating the commanded point.

Worst case they add directly, so size the cell on the linear sum; the root-sum-square figure predicts what a production run will see:

For a half-millimetre target, this allocation gives every stage a share it can beat and measure:

The four targets sum to 0.50 mm, the tolerance itself. Each stage has to come in under its target as measured, or the cell ships with no margin at all.

Stage 1: calibration error

Calibrate in place, at operating temperature, under the lighting you will run. The residual after the solve is your calibration error until drift moves it. Keep the mean under your allocation — 0.15 mm here — and redo any pose near twice the mean rather than averaging it away; camera calibration residual values: average and max targets explains how to read the report. Calibration work pays directly: a NIST study of vision-guided industrial robots cut median localization error by 97%, to 0.3 mm, through better use of the calibration points alone. If rotation results lean one way, the usual culprit is the quaternion convention — see the hand-eye calibration orientation error fix.

Stage 2: model error

The taught model is a promise about the part. Teach it at final part height so scale matches production; teach pattern matching at final part height shows why scale drifts otherwise. The part's own print matters more than any camera choice: a weld nut whose datum-to-seam distance varies ±0.3 mm on the drawing spends 0.3 mm of budget before the vision system powers up. Locate the feature the process actually uses — a machined hole, a coined mark — instead of a silhouette edge that moves with stamping tolerance. With a 0.05 mm model target, taught feature and real feature must agree within about a pixel.

Stage 3: locating error

Size the pixels first: pixel size at or under locating budget / 4. Subpixel edge detection resolves a fifth to a tenth of a pixel under decent contrast, so a four-pixel budget leaves the algorithm an order of margin and lets the mapping residuals already counted in calibration dominate. For a 0.20 mm locating target that is 0.05 mm/px; the camera resolution calculator turns your tolerance into the pixel number. Small parts also need field coverage for stable angle results — the 10% field rule gives the coverage-to-precision link. Contrast decides whether the subpixel math holds, so backlight the edge or run dark field on it. A compact global-shutter body keeps the capture pause short: the MV-CU020-90GC pairs a 2.5 MP sensor with GigE delivery in a C-mount, 1/1.7" format, one of many area scan cameras that fit a fixed locate.

Stage 4: picking error

Add the arm's published repeatability — typically 0.05 to 0.10 mm class on a mid-size six-axis arm — then everything the tool adds: jaw play, grip-force variation, part settling in the gripper. Measure rather than assume: let vision command the same reference point 30 times, re-gripping the part each cycle, and read the spread at the tool point with a dial indicator. The 3σ spread you measure is the real picking stage. If it returns 0.10 mm the stage holds its allocation; if it returns 0.25 mm, no camera setting recovers the budget.

Worked example: a spot-weld target at the drawing tolerance

Inputs: tolerance ±0.5 mm on a weld-nut placement, from the cell drawing; stage allocations 0.15, 0.05, 0.20 and 0.10 mm from the table above; an assumed camera of 1600 × 1200 pixels on a 1/2" sensor, 6.4 mm wide; working distance 200 mm, set by the welding gun's clearance.

Rule: the linear sum stays at or below the tolerance, and the pixel stays at or below a quarter of the locating allocation.

1. Pixel budget: 0.20 mm / 4 = 0.05 mm/px.
2. Field of view: 1600 px × 0.05 mm/px = 80 mm.
3. Magnification: m = 6.4 mm / 80 mm = 0.08.
4. Focal length: f = WD × m / (1 + m) = 200 mm × 0.08 / 1.08 = 14.8 mm, so the nearest standard lens is 16 mm.
5. Pixel size with the 16 mm lens at 200 mm: m = 16 / (200 - 16) = 0.087, FOV = 6.4 / 0.087 = 73.6 mm, pixel = 73.6 / 1600 = 0.046 mm/px.
6. Locating stage: 4 × 0.046 = 0.18 mm, under the 0.20 mm allocation.
7. Total: 0.15 + 0.05 + 0.18 + 0.10 = 0.48 mm worst case; across production, sqrt(0.0225 + 0.0025 + 0.0324 + 0.0100) = sqrt(0.0674) ≈ 0.26 mm.

Result: a 16 mm lens at 200 mm working distance over a 74 mm field holds the target with 0.02 mm of worst-case margin and about 0.24 mm of statistical margin. Decision: commission the cell when the measured mean residual stays at or under 0.15 mm and the re-grip spread at or under 0.10 mm. If the residual reads 0.20 mm or more, recalibrate in place with more poses before touching anything else — that stage returns the most millimetres per hour spent. The lens focal length calculator runs the same steps on your own sensor.

Fixed camera or wrist-mounted camera

A fixed overhead camera calibrates once, and its budget carries arm repeatability only inside the picking stage. The mount does the silent work: bolt it to its own pad, never to the robot base, which flexes with every move. A floor stand such as the LS-MNT-0040 — a 700 mm column on a 400 × 450 mm base — gives the stiffness and height travel a cell like this needs; the camera stands range adds slide stages for focus adjustment. The full walk-through is camera above the robot base: eye-to-hand calibration setup. A wrist-mounted camera rides the flange, sees the part at the point of use and shrinks model error, but every image inherits the flange pose, so calibration carries the arm's own error inside it; robot mounted camera calibration: capture and solve walks that path. Decision: fixed camera for fixture-based cells where the part presents in one zone; wrist camera when the part sits at unknown height or one camera serves several stations.

When the budget cannot close

Three failures have standard answers. If the part's own tolerance spends the budget, no camera helps: locate the datum the process uses and re-baseline against the drawing. If height or tilt is unknown, 2D cannot hold the budget, because scale error grows with ΔZ/WD — a 2 mm height change at 200 mm working distance shifts field scale 1%, or 0.8 mm across an 80 mm field, more than the entire drawing target. Move to 3D sensors: a projected structured-light camera such as the MV-DPS200P-02 returns the part's full pose in one capture, and the 3D camera working distance and positioning guide sets the mounting geometry. If one camera must serve conveyor and fixture work at once, 2D or 3D camera for conveyor pick and place maps the choice. The compact 3D option for pose at the point of use sits below.

Checks before you commit

• Mean calibration residual measured in place, under your allocation; redo any pose near twice the mean.
• Pixel size computed from the chosen lens and sensor, at or under a quarter of the locating allocation.
• Re-grip spread over 30 cycles, 3σ under the picking allocation.
• Teach feature equals the process datum, taught at part height.
• Reference-part check every shift; full recalibration after any collision or mount change.
• The four measured stages, summed, written on the sign-off sheet against the drawing tolerance.

Size the hardware from the camera finder and the stand finder, or start at the camera configurator and stand configurator. More applied setups live in the applications topic. If you want a second pair of eyes on the budget, send the drawing and your measured residuals through the contact page and we will check the allocation before you order.