Robot cell vision recalibration with one reference target

You run the same arm in a different cell every week, and the wrist camera keeps reporting part poses in the robot base frame as if nothing had moved. A planar locate calibration is only valid for one robot, one camera position and one work plane, so every relocation silently offsets every located pose. The camera-to-flange bolts did not move — only the cell did — which makes a full hand-eye re-run wasted effort most Mondays. This guide shows the faster path: one fixed reference target, one located pose, one planar shift applied to your pick frames, with the camera, lens and target arithmetic shown and the accuracy stated up front.

What the snapshot calibration actually locks in

A wrist-camera locate kit calibrates two relationships at once: the camera against the robot flange, and the work plane under a taught snapshot pose. Parts come back as poses in the robot base frame, and those poses stay truthful only while the robot, the camera position and the work plane are unchanged. Teach that first plane carefully when the cell is built — hanging robot vision calibration: teach the work plane covers that first teach — then treat every relocation as maintenance of one datum.

Moving the arm to another cell leaves the camera-to-flange transform intact but puts a different plane under the snapshot pose. Located poses then carry the full placement error: 2-3 mm with locating pins, tens of millimetres if the fixture was rebuilt by eye. Nothing in the numbers warns you; they look perfectly plausible.

Re-datum, do not re-teach. Measure where the cell datum went, shift your frames by that amount, and leave the published transform alone — the same philosophy as publishing eye-to-hand calibration as a static camera transform: one calibration underneath, offsets layered above it.

The method: one target, three points, one shift

Bolt a target with three circular fiducials of known spacing to the fixture that defines your datum. Record its located pose in the first cell. After each move, jog to the saved snapshot pose and locate the target again: the difference between taught and measured pose is a rigid planar transform — translation in X and Y plus rotation about Z — applied once to every taught pick pose.

Take theta from the direction change of one fiducial pair, t from the shift of the three-point centroid, then push every taught pick pose through the transform. The measurement is differential: lens distortion, sensor skew and mounting error sit identically in both images and cancel, so only real cell movement survives the subtraction.

If the kit publishes the located pose

Many locate kits expose the found pose in the robot base frame. Register the plate in the kit as though it were any other part, and that reported pose doubles as your datum: after each relocation, the difference between taught and measured becomes a user-frame offset. Same math, no extra tooling. Respect the planar assumption: the reported height belongs to the taught plane, and a relocation that tilts the plane never shows up in the output — the witness fiducial below is the only guard against it.

Where the height term comes from

A planar locate gives no trustworthy Z. Two sources work. Clamp the fixture at a known height with hard stops and keep the taught Z. Or read the fiducial spacing: the robot holds the camera at the saved pose, so a spacing change is a distance change — Δz ≈ z × (s_measured − s_taught) / s_taught. Plate tilt enters as its square, about 0.01 mm of false Z at 0.5°, so plate flatness matters more than levelling.

If the camera stays in the cell

Eye-to-hand reverses the problem: camera and fixture stay, the robot rolls in, and the lost relationship is cell-to-robot-base. Locate the same target with the fixed camera, then touch one known point with the tool tip or run a four-point base teach. Mount the camera on a slide-table camera stand so it re-seats over a floor mark after every move:

The LS-MNT-0040 is a professional slide-table type stand, height 700 mm, 400 x 450 mm base. Its XY table is the fine adjustment that puts the camera back over the mark.

Re-datum procedure after every relocation

1. Make the target: a matte white plate with three black circles 6 mm across at the corners of a scalene triangle, longest side 150 mm, plus a fourth circle outside the triangle as a witness. Flat within 0.05 mm across the span. Caliper-measured geometry is fine — the method is differential, so absolute errors cancel between visits and only damage shows.
2. Fix the plate to the datum fixture with two dowels and a screw. Two dowels fix rotation; tape and magnets do not.
3. In the first cell, teach the snapshot pose and work plane with the kit wizard, then record the target's located pose — X, Y, rotation — next to the taught pick poses.
4. Relocate the arm and seat the fixture against its pins. ±20 mm of slop is fine; the field of view absorbs it.
5. Clear the cell, keep the operator outside the envelope with the arm in reduced-speed mode, and jog to the saved snapshot pose. The encoders reproduce the taught pose — the pose is not the unknown, the cell is.
6. Locate the target and read X, Y and rotation; compute theta, t and Δz from the spacing change.
7. Apply the result as one user-frame offset to every taught pick pose. Never re-teach picks by hand after a move.
8. Check the witness residual, then make the first approach 10 mm above the taught pickup before committing to contact.
9. Log the shift. More than 25 mm of translation or 0.5° of rotation between visits means the cell was rebuilt wrong — fix the fixture instead of absorbing the error.

Sizing the camera, lens and target

Set the targets first: hold ±0.1 mm in X and Y, ±0.05° in rotation, absorb ±20 mm of slop around a 150 mm fiducial baseline. The field of view must then be 150 + 2 × 20 + 10 mm of margin = 200 mm wide and about 150 mm high, with the taught target centred — slop walks fiducials toward the frame edges, and edges are where distortion lives (the calibration target field of view and lens coverage margins guide quantifies that margin).

For the image, take a 12 MP area scan camera. The LS-CMR-0166 is a 12 MP USB3 camera, IMX226, 1/1.7" rolling shutter, 30 fps, 4000 × 3000 resolution. Rolling shutter costs nothing here because the arm stands still for the snapshot.

A 1/1.7-inch sensor has an active area of roughly 7.4 × 5.5 mm. Across a 200 mm field of view, 4000 pixels work out to 0.05 mm per pixel in the cell plane, and each pixel is about 1.85 µm on the sensor (7.4 ÷ 4000).

m = 7.4 ÷ 200 = 0.037. From a working distance of 336 mm: f = 336 × 0.037 ÷ 1.037 = 12.0 mm — a standard 12 mm C-mount lens. Check it from the lens side: m = 12 ÷ (336 − 12) = 0.037, field of view = 7.4 ÷ 0.037 = 200 × 150 mm. The target fills the frame with room left for the slop.

A 6 mm fiducial spans 6 ÷ 0.05 = 120 pixels. The centroid of a filled circle that size repeats to about a tenth of a pixel, which is ±0.005 mm in the cell frame — twenty times finer than the ±0.1 mm budget.

Both interfaces carry the job; pick by cable run across the cell:

The GigE version of the same 12 MP IMX226 sensor streams at 9 fps, and that is plenty because the snapshot is static. Light the plate flat and matte: a ring light 20-30° off axis, exposure 5-10 ms, background near mid-grey at about 120 of 255. A short exposure keeps servo dither out of the edges even with the arm parked.

The accuracy you can hold

Vision-only arithmetic: rotation from a 150 mm baseline with 0.005 mm centroids is 0.005 ÷ 150 = 3.3 × 10⁻⁵ rad = 0.002°. Translation, the mean of three centroids, stays under 0.01 mm. Height: two centroids carry about 0.007 mm of pair noise, so Δz resolves to 336 × 0.007 ÷ 150 = 0.016 mm before plate flatness eats into it.

The cancellation has one limit: it is exact only while the fiducials land on the same pixels. Slop walks them across distortion — a lens with 0.5% distortion shifts a fiducial 0.1 mm per 20 mm of walk, while a low-distortion lens at 0.1% keeps the same walk at 0.02 mm. That is what the locating pins buy you.

Neither vision term dominates; the arm does. Published pose repeatability for mid-size collaborative arms runs ±0.03 to ±0.1 mm, and the datum inherits it because the camera rides the wrist. Commit to ±0.1 mm in X and Y, ±0.05° in rotation and ±0.05 mm in Z per re-datum, and you meet that with margin.

Reading the kit's own locate output instead of running your own tool inherits the kit's pose resolution; plan on ±0.2 mm and lean on the witness. Your own three-circle centroid — the robot vision for locating part center points approach applied to circles instead of parts — delivers the arithmetic above.

Check the re-datum before the first pick

• Witness residual under 0.1 mm. The fourth fiducial takes no part in the fit, so its residual is a free, honest grade of the shift.
• Repeat the locate three times without moving: spread under 0.02 mm in X and Y. More means lighting or threshold drift, not cell movement.
• Confirm Z once per cell with a gauge block or paper feeler under the gripper; trust the spacing method after it agrees twice.
• First pick at reduced speed, 10 mm above the part, then contact.
• If picks still land off after a clean re-datum, walk the diagnose an eye in hand pick point offset sequence to separate camera, calibration and tool errors.

Mistakes that move the datum

• Symmetric fiducial layout. Three circles on an equilateral or isosceles triangle can match rotated 180°; make no two spacings equal.
• Target on the robot's table instead of the datum fixture. The datum must ride with whatever defines the work.
• Absorbing a big shift silently. The math will happily apply 60 mm; beyond 25 mm the cell was rebuilt wrong.
• Touching the bracket or focus between the teach and the re-datum. That changes camera-to-flange, and only a fresh wizard pass fixes it — the vision calibration recovery after camera or robot changes sequence lists what to re-teach.
• A glossy plate under a hard point-source light. Specular edges wreck centroids; matte finishes win.

When one reference target is not enough

Three cases break the method. Changed camera hardware — bracket moved, lens refocused or swapped — kills the differential basis, so run the full wizard again; the robot hand-eye calibration: how nine-point alignment works procedure is the from-scratch path. A work plane that can tilt more than about 0.5° between builds thins the planar assumption: add a second target at a different height or move to a full hand-eye solve. And a joint replacement or lost mastering takes the robot itself out of truth — that is a robot remastering and frame setup job before any vision matters.

If relocations become a weekly routine, budget honestly: the vision re-datum takes minutes, the safety re-validation of a moved cell takes the day — how vision and safety work size a robot cell project breaks that split down.

If the kit is sealed, with a fixed lens and no access for your own tool, swap the camera side for a C-mount smart camera and run the fiducial tool on-device:

The LS-SMC-0017 is an 8.8 MP smart camera with a 1" sensor, C-mount and global shutter. Run the same 12 mm C-mount lens on it and the three-point math executes on-device with no separate PC.

Relocation checklist

• Reference plate dowelled to the datum fixture, not the table.
• Witness residual under 0.1 mm, logged with the shift.
• Shift under 25 mm and 0.5°; larger means fix the fixture, not the offsets.
• Z confirmed once per cell with a gauge.
• Camera bracket and focus untouched since the last wizard pass.
• First pick at reduced speed with the 10 mm standoff.

Size the camera, lens and light for your own target span in the system configurator, or send the target drawing and shift log through the contact page and we will check the arithmetic with you.