Wrist camera mounting orientation: fix hand-eye calibration poses

A hand-eye calibration that centers perfectly and then sends the robot to a first pose where the calibration board is cropped or missing is, in practice, a mounting fault. The usual culprit is roll: the camera body is rotated about its optical axis relative to the robot flange, most often because a custom bracket or adapter plate leaves the dowel pin misaligned. This guide gives you the symptom pattern, two image-based tests that measure the roll, the framing arithmetic that sets how much roll you can tolerate, and the keying fix that makes the error impossible to repeat.

Symptom: the poses compute, then pose 1 loses the board

The sequence is distinctive. Centering works: you jog the wrist until the board sits on the crosshair and the tool accepts it. The solver computes its pose list without complaint. The robot drives to pose 1, and the board is skewed in a corner, half out of frame, or absent altogether; after the tool's timeout it reports that the board is not visible.

The discriminator is repeatability across scene changes. Move the board, move the snapshot position, change the lighting — the same bad pose 1 comes back. A scene fault moves when the board moves; a geometry fault moves with the camera. When the same wrong pose survives every scene change, the camera-to-flange relationship is the suspect.

Pose selection guides such as eye-in-hand calibration pose selection: robot positions that work all assume one thing before they start: a camera whose image axes follow the flange axes. That assumption is exactly what a rolled mount breaks.

First suspect: the camera is rolled on its mount

Roll is rotation of the camera body about its own optical axis — image up pointing somewhere other than where the flange says up. It leaves the optical axis itself unchanged, which is why centering still passes: centering aligns the axis, not the in-plane axes. The pose planner, however, predicts how each wrist motion will move the board across the image using the assumed camera frame. Roll the camera by an angle θ and every one of those predictions rotates by θ, so the first commanded tilt moves the board along the short axis of the frame instead of the long one.

Factory flange mounts key this with a dowel pin: the pin fixes the clocking, the bolts only clamp. The failure appears when the camera mounts through something else — an adapter plate, a printed shell, a bracket drilled without a pin — where the pin misses its hole or the camera bolts through clearance holes at whatever angle the assembler chose.

The same applies to a wrist-mounted 3D sensor: a stereo or structured-light head rolled on its bracket displaces every captured cloud the same way, so the checks below are worth two minutes either way.

Building the wrist camera from a board camera

Most wrist builds start from a compact board-level camera bolted close to the flange. A board camera carries no dowel of its own: the drawing marks one edge as image-up, and the bracket has to hold that edge where the flange expects it. Check the drawing before you design the plate, and keep the hole pattern square to the sensor.

Both of these mount through an M12 lens seat, so the lens is a separate choice — the squareness comes from your plate, not from the barrel.

Test 1: measure the roll from one image

Put the wrist in a known pose: flange face parallel to the bench, board edge parallel to the bench edge, camera looking at the board. Capture one image and pick two points along the board's top edge in pixels. The roll is the slope angle: θ = arctan(Δy / Δx).

An edge running from pixel (200, 400) to (1800, 520) gives Δx = 1600 px, Δy = 120 px, and arctan(120 / 1600) = 4.3°. A 4.3° roll is enough to skew every capture in the dataset. Gross clocking is even plainer: a board standing on end in a landscape frame is a 90° answer.

Cross-check with the swap test: jog the wrist ten degrees in pan only. With a square camera the board slides horizontally in the image. If it slides vertically instead, the camera is rolled near 90°; a diagonal slide means a roll in between. Two minutes, no solver run, no remounting.

Test 2: how much roll the framing can absorb

FOV_w ≈ sensor_w × WD / f
margin = (FOV_w − board_w) / 2
edge crop when: offset × sin θ + board_w / 2 > FOV_h / 2

Example inputs: a 1/2.9-inch sensor (1440 × 1080 pixels at 3.45 µm pitch → 4.97 mm × 3.73 mm), a 6 mm lens, a first-pose distance of 300 mm, and a 150 mm square board. Then FOV_w ≈ 4.97 × 300 / 6 ≈ 248 mm and FOV_h ≈ 3.73 × 300 / 6 ≈ 186 mm, so the board leaves a margin of (248 − 150) / 2 = 49 mm left and right but only (186 − 150) / 2 = 18 mm top and bottom. That asymmetry is what turns a clocking error into a crop.

Assume the wizard's pose 1 holds the board 35 mm off the image centre horizontally, inside the 49 mm margin. Roll rotates that offset by θ, and the vertical component decides the outcome:

Pose 1 alone survives a roll of asin(18 / 35) ≈ 31°; the wizard's later poses tilt further and eat the margin sooner, so a rig that barely passes pose 1 fails again one board-move later. The working limit is a few degrees, and the target — achieved by pinning — is under 1°.

Rerun the first line with your own sensor width and focal length:

Fix: key the camera to the flange, then recapture

1. Establish the intended clocking. The camera drawing marks the image-up edge, and the hand-eye wizard's setup drawing shows which flange direction it must face. On a factory flange mount the dowel already fixes the clocking, so the misalignment lives in the adapter between bracket and camera.
2. Refit with both dowel holes engaged. A pin in one hole still lets the body rock a few degrees; bolts through clearance holes alone let it sit at any angle. On a custom bracket, ream the pin holes to the dowel size and drive proper alignment dowels — the pin locates, the bolts clamp.
3. Scribe a clocking mark across the camera-to-bracket joint so the next remount cannot clock silently.
4. Re-run Test 1 before any solver run. Under 1° measured roll, proceed. At 4° or more, fix the metal first: recalculating on a rolled mount produces a confident wrong answer.
5. Recapture the dataset and recalculate. The poses are computed from the captured images, so an old dataset stays wrong no matter what the hardware now does; the hand-eye calibration pose capture workflow covers the capture order.
6. Store the reference image and the measured angle with the calibration record, so the next engineer can tell a mount shift from a software change in minutes.

Before you rerun the solver:

• measured roll under 1°
• both dowels engaged, clocking marks aligned
• mount screws tight and nothing moved since the snapshot
• board is the one specified for your robot model and fully inside the frame at snapshot distance

If the mount is square and pose 1 still fails

Ordered by how often each turns out to be the cause:

1. The mount moved between the snapshot and the capture. A loosened screw, a cable levering on the housing, one bump — every pose computed afterwards inherits the shift. Re-torque, re-center, recapture. Any remount or tool change invalidates the calibration; wrist camera tool change: redo the hand-eye calibration covers the tool-change case.
2. The board does not match the robot model. Calibration boards differ in size and pattern between models, and the wizard's poses assume the expected one. A board that is too small crops at the margins computed above; one that is too big overflows at every pose.
3. A camera-software update changed the routine. If the failure started right after an update, read the release notes and recapture with the new version rather than reusing a stored dataset. Chasing an older installer is not a fix — take the current release from the vendor's support channel. Robot vendors' hand-eye tools each define their own snapshot-and-pose conventions, and a worked hand-eye calibration walkthrough for a collaborative robot shows how much the pixel-to-robot transform depends on getting those conventions right.
4. Intrinsics went stale. If focus or aperture was turned after the intrinsic calibration, corner detection degrades and the solver fails in ways that look like pose faults; the camera intrinsics for hand-eye calibration guide covers those first.

More system-side setup guides sit on the systems topic page.

Prevention: make the clocking impossible to lose

Pin or mark every custom bracket at design time: a reamed dowel in both holes costs less than one failed calibration shift, and an engraved clocking mark survives wiping and coolant better than memory. Re-run the roll test after every remount, not only after the first install.

The measurement needs the camera still for two exposures, which a bench rig makes trivial. A slide-table camera stand such as the LS-MNT-0040 holds the camera and board square while you take the pair, and the same rig serves lens and focus checks later.

Mechanical detail decides whether the calibration survives life on the wrist: stiffness, flange fit and vibration are covered in robot wrist camera mounts: stiffness, flange fit and vibration.

Next step: if you are speccing a new wrist camera end, start at the camera finder or the camera step of the configurator; for brackets and bench rigs, the accessories finder lists what the stand category carries. If a calibration keeps failing after the checks above, contact us with the reference image and the measured roll angle.