Vision guided picking: fixing a camera calibration warning

A picking cell that raises a "camera lens is not calibrated" warning almost never has a broken camera. The station simply has no valid pixel-to-robot transform loaded, and most of the time the calibration run never captured enough of the target to build one — the sheet sat partly outside the field of view. The vendor checklist for this warning on one widely used picking platform carries exactly two items: the calibration paper must be in view, and the robot name must not start with a digit. The sections below work the causes in order of likelihood, size the coverage with real numbers, and end with a repeatability test that proves the fix held.

What the warning actually means

The station stores one transform per camera: a mapping from image pixels to robot base coordinates, valid for one lens, one focus setting and one mount height. The warning fires at load time when nothing usable is there. Three causes cover nearly every case, in this order: the calibration run captured no target features, the wizard could not bind the robot, or the calibration ran but was never saved — or has since been made stale by a lens or mount change. Some controllers go further and block the pick routine outright until calibrated camera coordinates exist, as picking routines that block on uncalibrated camera coordinates shows on one commercial cell.

If you are still choosing the guidance architecture rather than repairing one, read 2D vision pick and place: selecting a planar guidance system first.

Check the target coverage first

The hard rule: every feature of the sheet must be inside the image when the wizard grabs. Corners clip first — a lens whose image circle is smaller than the sensor rounds the frame's corners off, and a clipped feature is a dropped feature. Beyond that, spread the target. Aim for it to span one-third to two-thirds of the frame width, with its outer features at least a tenth of the width clear of the edges. A sheet smaller than about a quarter of the width forces the fit to extrapolate into tray corners it never measured, and pick points out there drift.

The sheet must also be flat at pick height: paper on a surface plate is fine, paper on a wavy tray rim is not — the tolerances are in camera calibration target flatness: when paper is enough. Sizing the target against the lens's coverage has its own arithmetic in calibration target field of view and lens coverage margins. If the station reports a missing target frame rather than a missing calibration, start at hand-eye calibration target frame missing: what to check.

Size the field of view with numbers

Most failed reruns fail again because the sheet is centred and visible but tiny in the frame. Do the arithmetic once, then rerun.

Worked example: an overhead camera looks at a 480 × 340 mm tray with 30 mm of margin wanted each side, so the field of view must be at least 540 × 400 mm. The camera sits 700 mm above the tray. A 1-inch format sensor is nominally 12.8 × 9.6 mm, a 16.0 mm diagonal.

• Magnification: m = 12.8 / 540 = 0.024
• Focal length: f = 700 × 0.024 / 1.024 = 16.4 mm, so the standard 16 mm lens
• Delivered coverage: FOV = 700 × 12.8 / 16 = 560 mm wide, 700 × 9.6 / 16 = 420 mm tall, about 20 mm spare beyond the margin on each axis

The camera for this station is the MV-CS200-10UM, a 20 MP mono USB3 camera on a 1-inch C-mount sensor with a rolling shutter. Twenty megapixels at a 3:2 aspect works out to about 5472 × 3648 pixels, so across the 560 mm field each pixel is 560 / 5472 ≈ 0.10 mm. The lens is the LS-FFL-0567: 16 mm focal length, C-mount, built for 1.1-inch sensors, with a 16-22 mm image circle and a 20 MP resolution grade. The sensor's 16.0 mm diagonal sits inside that circle. The stand is the LS-MNT-0040, a slide-table camera stand of 700 mm height on a 400 x 450 mm base. That class of slide-table stand gives the fine height trim the calculation asks for.

Now check the target against those numbers: a 200 mm sheet spans 200 / 0.10 = 2000 pixels, 36 percent of the width — inside the band. The smallest feature you pick, say 3 mm, covers 30 pixels; five pixels is the floor for locating a pick point. Set the aperture near f/5.6. That is mid-range on the lens's F2.8-22 scale, and at this magnification the depth of field runs about ±100 mm, so a sheet a few millimetres off pick height changes nothing.

To swap any part of the station: the area scan cameras finder filters by sensor format, mount and shutter; the machine vision lenses finder by focal-length band; the camera stands range by stage type. Once the focal length is settled, the lens finder is the quickest path.

The robot name that stops the wizard

Second on the list, and the cheapest to fix. The published checklist for this warning names two items and this is the second: the robot name must not begin with a number. The trap is that the name looks healthy in the network monitor, which prints the raw string, while the wizard parses it — a leading digit is read as a number, the station never binds, and the calibration has no robot to attach to. Rename with a letter first, 2Cell becomes Cell2, then reopen the wizard. Use letter-first names for every cell, camera and station you commission; it removes a whole class of silent parsing errors for free.

Did the calibration actually save?

A completed wizard is not a saved calibration. The station must write the transform and load it back. The usual failure points: the cell lost power during the save; two stations share one camera and only one holds the transform; the camera was replaced, even with an identical model, after the original run; or the focus ring crept since commissioning, quietly shifting the transform while images still look sharp. Open the station, read back which calibration it loads, and restart once to confirm the warning stays gone. A cell that sits in sunlight can do this slowly across a shift; the camera focus drift with temperature diagnostic separates thermal drift from a mount that slipped.

Prove it: the ten-point repeatability test

1. Lay the calibration sheet flat on the tray at pick height and let vision report one marked point.
2. Command the gripper tip to that point without closing on a part, at reduced speed and in manual mode. The first motion after any recalibration is always supervised.
3. Mark where the tip lands, then repeat ten times.
4. Measure the spread of the ten marks. At 0.10 mm per pixel, a rigid mount holds the marks inside about 0.5 mm; vision alone contributes 0.1-0.2 mm and the robot and mount make up the rest.
5. Rotate the sheet 90 degrees and repeat. A spread that grows with rotation points to an unflat target or an unlevel tray, not to the calibration.

The verdict: under 0.5 mm, the calibration is good. Between 0.5 and 1 mm, recalibrate and repeat the test. Over 1 mm, stop and fix the mechanics first. Check the fit itself as well as the picks — camera calibration residual values: average and max targets sets pass levels for the residuals. The same check on an integrated platform is walked through in verifying camera calibration with RMSE targets.

Once the transform holds, robot vision for locating part center points covers the pick-point extraction that follows it.

When recalibration will not fix it

Two cases look like calibration faults and are not. The first is height. A planar transform maps pixels onto one plane at a known height; stacked parts, cupped trays and tilted pallets put the true pick point above or below that plane, and no rerun of the wizard recovers it. Take Z from a spacer stack, the part model or a distance sensor. How far one plane can be pushed is shown in measure planar distances with a calibrated camera. The same limit on a real cell appears in plate picking where Z must be handled outside the image.

The second is motion. The example camera carries a rolling shutter, right for a tray that stops under it. Pick from a moving conveyor instead and the rolling shutter shears the geometry as the part crosses the field, so the transform is applied to a distorted image. Use a global-shutter camera, or strobe the part to freeze it — set the shutter filter first in the camera finder. Placement of the camera along the line has its own rules in conveyor tracking camera placement relative to the robot.

Checklist before you rerun the wizard

• Sheet flat on a plate, fully in view, spanning one-third to two-thirds of the frame width.
• Sheet at pick height, not propped on the tray rim.
• Lens rated for the sensor format or larger; the 16.0 mm diagonal needs at least that much image circle.
• Focus sharp on the sheet's finest feature, focus ring locked.
• Aperture set and recorded, exposure fixed rather than automatic.
• Robot, cell and station names all start with a letter.
• One station open, one camera, the right pair; save and read back before closing.
• Restart the station software; the warning must stay gone.
• Ten-point test at reduced speed; spread inside 0.5 mm.
• One photo of the rig — working distance, focus mark, aperture, sheet position — filed with the station records.

If all three checks pass and the warning survives, send us a grab of the frame and the warning text through the contact page, or rebuild the station in the system configurator with the numbers above.