Restore vision accuracy after a robot joint replacement

Restore the cell in three moves, in this order: recalibrate the robot — the single-joint quick method (~30 min) when picks tolerate ±3 mm or looser, the plate calibration (60–90 min) or the two-robot run for the 0.1–0.5 mm class — then re-teach the vision snapshot positions in the camera-locate nodes and save the installation file, then verify the pixel-to-robot scale within 1% of the pre-service log with a 200.0 mm dot pitch.

A joint swap on a vision-guided robot usually arrives disguised as a vision fault: picks drift, the offset grows toward one corner of the table, and the camera gets the blame. The camera is almost never the problem. The replaced joint moved the robot's own kinematic datums, so the transform between pixels and robot coordinates is now wrong by millimetres, and re-teaching on the vision side alone will not hold it. This guide walks the recovery in the order that works, with the tolerances, pixel counts and time budgets for each path.

Can you run the vision system after a robot joint replacement?

Yes. The camera, lens and stored calibration data are untouched by the swap; what moved is the robot-side datum those numbers assumed. Recalibrate the robot, re-teach the snapshot positions, save the installation, and the cell returns to its pre-service accuracy in one to three hours of cell time.

A joint replacement is a mechanical repair with a mandatory calibration step at its end — the service manual ends there because the new joint's zero mark never lands where the old one's did. The vision system does not need rebuilding; it needs the robot underneath it to tell the truth again. A few millimetres of pick offset, worse in some poses than others, is the signature of a kinematic datum shift. When the same signature appears with no service done, the diagnosis path in robot camera calibration lost after restart: diagnose the shift applies instead.

Why does a replaced joint move the vision picks?

Because the new joint's zero position lands a fraction of a degree off, and 0.1° of datum error puts the tool point 1.0 mm out at 600 mm of reach — before the vision system adds any error of its own.

The controller computes the flange pose by integrating joint angles through a set of calibrated link offsets. Factory calibration measured those offsets to hundredths of a degree; the replacement joint brings an unverified datum, and every untaught pose downstream inherits the error. Two properties decide how the cell behaves after the swap:

• Repeatability survives. Returning to a taught pose depends on the mechanics, which are new and tight; published repeatability for this class of arm stays at a few hundredths of a millimetre.
• Absolute accuracy breaks. Computing where an untaught pose sits depends on the datums, and every vision-guided pick is exactly that: a robot coordinate calculated from pixels.

That split is why the cell looks broken while manual jogging looks fine. The error is a small offset plus a small rotation that turns with arm pose, so a constant correction taught at one point works there and misses by a different amount somewhere else — and the miss grows with reach.

Which robot recalibration method should you use?

Use the single-joint quick method when picks tolerate ±3 mm or looser: about 30 minutes of work that leaves 1–5 mm of absolute error. For anything tighter, use the plate method or the full two-robot calibration — both restore pre-service accuracy, and the two-robot run itself takes under an hour.

The steps of each method are in the robot's calibration manual — the tier choice below is yours, the procedure is not.

The quick method leaves millimetres on the table because it re-zeros the swapped axis only; the factory-measured offsets that couple it to the other axes stay out of the loop. The plate method brings them back with a calibrated plate jig and one robot — budget 60–90 minutes including fixturing. The two-robot method pairs the serviced arm with a healthy one and a calibration kit; the run itself takes under an hour and restores the calibration the cell ran on before.

Decision, case by case:

• ±3 mm or looser, one fixed pick point, loose presentation: single-joint quick method. Accept that the offset varies across the workspace.
• ±2 mm: plate calibration. The quick method is a gamble at this tolerance; the plate run is the safe hour.
• ±1 mm or better, corrections applied across the table, or a wrist camera: plate calibration with one robot, or the two-robot method when a second arm is on site and the vision application must run exactly as before.

Does the hand-eye calibration survive the joint swap?

Usually, yes — after a full calibration. A plate or two-robot run restores the kinematics the hand-eye transform was measured against, so the stored camera-to-flange or camera-to-base matrix stays valid; after the single-joint quick method, plan on re-measuring it.

The hand-eye transform — camera to flange on a wrist camera, camera to base on a fixed one — describes geometry, not joint datums. A full recalibration puts the robot back where that geometry was measured, so the matrix itself holds. The single-joint method does not fully do that; it leaves cross-coupled error behind.

Check it without a full re-run: fix a reference target to the table, drive to a taught inspection pose, and compare the detected target position with the pre-service log. Beyond a third of the pick tolerance — 0.7 mm in a ±2 mm cell — re-run the hand-eye calibration with a fresh pose set; eye in hand calibration pose selection: robot positions that work covers which poses make the new calibration stick. If more than the joint changed — camera moved, cable re-routed — the wider path is vision calibration recovery after camera or robot changes.

Do you have to re-teach the vision snapshot positions?

Yes. Snapshot positions are taught wrist poses, and the recalibrated robot places those joint targets at slightly different camera framings. Define new snapshot positions in the camera-locate nodes, then save the installation file — an unsaved teach is the most common failure in this recovery.

Two reasons the old poses are not good enough. The same joint targets now place the camera at a slightly different point, so the part sits off image centre, closer to the edge where lens distortion is stronger and measurements degrade. And exposure and focus were tuned for the old framing.

Re-teach in this order: define the snapshot positions again in the camera-locate nodes, re-check exposure on a real part, then re-teach pick reference frames stored in image coordinates. The coverage pattern in multiple wrist camera snapshot positions to cover a large table applies to the new poses exactly as it did to the old ones.

How do you verify the pixel-to-robot scale after service?

Measure it directly: put a 200.0 mm dot pitch or a precision ruler in view, read the pixel distance, divide, and compare with the pre-service log. The scale must land within 1% of the logged figure, and corner-and-centre offsets must stay under a third of your pick tolerance.

Worked numbers, overhead pick camera. A 1/1.8-inch sensor is about 7.2 x 5.4 mm; at 3072 x 2048 pixels that is a 2.3 µm pixel. You want 420 mm across the belt from a 700 mm working distance:

m = 7.2 / 420 = 0.017 f = 700 x 0.017 / 1.017 = 11.8 mm, so a 12 mm lens FOV = 7.2 x 700 / 12 = 420 mm scale = 420 / 3072 = 0.137 mm/pixel

Then the scale check, in pixels:

• A 200.0 mm dot pitch reads 1463 pixels apart: 200.0 / 1463 = 0.1367 mm/pixel. The pre-service log said 0.1358 — drift of 0.7%, inside the 1% limit. Pass.
• A reading of 0.145 mm/pixel — 6% out — is not a robot problem. The camera or lens moved; check the focus ring and the mount before touching the robot again.
• Offset test: drive the tool point onto a dot centre; a 3-pixel vision offset is 3 x 0.137 = 0.41 mm — pass. A single fixed target is all this needs; robot cell vision recalibration with one reference target covers the method.
• Repeat at four corners and the centre. Failures that rotate across those five points are residual kinematic error: move up one calibration tier.

Your own sensor and distance go through the same arithmetic here:

Should the camera move from the wrist to a fixed mount?

If this arm eats joints — high duty cycle, collisions, repeat service — yes. A fixed overhead mount keeps the camera-to-base transform through joint swaps, removes snapshot positions from the recovery entirely, and leaves only the robot recalibration to redo each time.

The trade is view flexibility against recovery cost. A wrist camera frames each part close up, but every joint service invalidates its hand-eye calibration and every snapshot pose. A fixed camera frames the whole workspace at once — size it with the same focal-length arithmetic above — and after a joint swap you recalibrate the robot and pick up production, with nothing to re-teach on the vision side. For the guidance background, 2D vision pick and place: selecting a planar guidance system covers the fixed-camera pattern end to end.

The parts are ordinary. Camera stands put the camera out of the arm's reach; the LS-MNT-0040 is a professional slide-table machine vision stand with a 700 mm height and a 400 x 450 mm base.

Among area scan cameras, a GigE mono model with pixels to spare for the worked example above is the LS-CMR-0053 — C-mount, 1/1.7" sensor, 12 MP resolution.

Vision controllers run it: the MV-VC3301H-128G66 carries 5+ LAN ports and 4 built-in light channels, enough for several GigE cameras and the lighting from one box.

The camera finder filters by sensor format and mount, and the configurator turns field of view and tolerance into a camera and lens.

What checks close the job before the cell runs again?

Ten consecutive logged picks at three workspace locations, every reading inside a third of your pick tolerance. It is about 15 minutes of testing that catches a half-done calibration before it ships bad parts — run it after every path, including the two-robot one.

• Pixel-to-robot scale within 1% of the pre-service log, measured at the working distance the cell runs at.
• Offset at four corners and centre under a third of the pick tolerance.
• Ten logged picks at three workspace locations, every one inside tolerance.
• Snapshot positions re-taught, exposure re-checked on a real part, installation file saved.
• Reference target reading filed with the service record — it is the baseline for the next recovery.
• Backups of the calibration and installation files taken after sign-off.

Keep the log with the service record; the next joint swap starts from numbers instead of symptoms. More vision system guides sit under the same topic, and if your tolerance or workspace does not fit the numbers here, tell the applications team the pick tolerance and table size, and the recovery plan gets written before the mechanic books the service window.