A robot camera calibration that stops partway with no error message is almost always a reach problem on a compact arm, not a lighting problem. The wizard computes its 18 calibration poses from the board it detects; when one of those poses sits outside the working envelope, the inverse-kinematics request for it returns false and the run aborts silently. The fix is mechanical, not photographic: print the grid smaller, 8.24 mm squares at a 58.8% scale on a UR3-class arm, and check the squares with calipers. This guide shows where to read the failure, how to separate reach faults from detection faults, and the numbers that put every pose back inside the envelope.
What the failure looks like
The sequence is distinctive. You press Calibrate, the robot moves to a few locations and captures images — a different count depending on where it started — then the image pane fills with the numbered pose grid, the camera preview flashes on and off, and the Calibrate button becomes available again. No error appears on the pendant.
The abort happens before the full pose set is captured, and two faults produce this exact silence: the camera lost the grid mid-run (a detection failure), or the arm could not physically reach one of the computed poses (a kinematics failure). The wizard's screen does not separate them. The controller log does, in one line.
Read the log before you change anything
The calibration routine writes each computed pose and its inverse-kinematics outcome to the controller log on the teach pendant. Open the log, run the wizard once, and find the pose whose inverse-kinematics request returns false. In the case tracked here, the log read the same on every attempt: pose 11 of 18, inverse kinematics false, calibration aborted.
| What the log shows | Cause | First move |
|---|---|---|
| The same pose number returns false every run, from every start position | The computed pose is unreachable | Scale the grid down |
| The failure point moves; the last captured image is blurry or washed out | The grid was lost during capture | Fix focus and reflections |
| The run stops right after centering, with no pose attempts logged | The board was never detected | Check standoff, glare and flatness |
A failing pose number that repeats across runs and across start positions is the signature of a geometry fault. If the captured-image count changes and different poses fail each time, work the detection path instead — the diagnosis and fixes in chessboard not detected in calibration cover it.
The causes, in order of likelihood
- An unreachable computed pose. The likeliest fault on a compact arm carrying a full-size grid. A board sized for a longer-reach robot leaves the arm extended after the centering step, and the remaining poses cross the working envelope or the wrist's joint limits. Test: the log's false line at the same pose on every run.
- The grid lost during capture. Focus drift or a specular hotspot on the board. Test: the last captured image in the wizard looks soft or glaring. Fix: matte reprint, remove the reflection, hold the standoff.
- A shadow crossing the grid. Usually the arm's or the gripper's own shadow, arriving with a late pose. Test: watch the board through the whole run. Fix: move the fill light or shift the board mount.
None of the three is fixed by moving the start position. The snapshot pose the wizard uses as its reference changes the approach path, not the pose set the board generates.
Why a big grid pushes poses out of reach
The wizard scales its pose pattern to the grid it finds. A larger grid needs a wider field of view, so the camera must stand farther from the board, and every computed pose inherits that standoff. On a 500 mm-reach arm already stretched after centering, the far poses land outside the envelope and the solver answers false. Smaller squares pull the whole pose set back inside.
This is the opposite constraint from hand-eye calibration target size: how big a board your poses need, which sizes the board for accuracy at a given pose distance. Here the board is sized so the poses stay reachable; accuracy follows, because the smaller squares still resolve comfortably, as the numbers below show. The camera kit's own setup pages put the camera 20 cm to 80 cm from the board, and the scaled grid keeps its poses in the near half of that window. Offline programming enforces the same rule from the other side — process simulation only accepts reachable robot poses — so a pose set that fails on the arm fails there too.
The compact-arm print is 58.8% of the artwork, and its final dimension belongs to your calipers, not the printer.
Worked example: an 8.24 mm grid on a 500 mm-reach arm
Inputs, each with its source:
- 18 poses computed per run, by the wizard's design.
- Failing pose: 11, inverse kinematics false, identical across four start positions (controller log).
- Arm: UR3-class.
- Required square: 8.24 mm, the grid size specified for compact arms.
- Print scale: 58.8% of the full-size artwork.
- Camera-to-board window: 20 cm to 80 cm (the kit's setup pages).
Rule: scale the artwork until a printed square measures 8.24 mm, then let the measurement decide.
Arithmetic. The full-size square is 8.24 mm ÷ 0.588 ≈ 14.0 mm, and 14.0 mm × 0.588 = 8.23 mm off the printer — three hundredths short of target, which is exactly why the measured square, not the scale number, is the requirement.
Detection margin, with a camera typical of these kits — illustrative inputs: a sensor 5.6 mm wide with 1440 pixels across (3.9 µm pixels) behind an 8 mm lens:
An 8.24 mm square then covers 8.24 × 8.2 ≈ 68 px, inside a 5.6 × 250 ÷ 8 = 175 mm field of view; a ten-square span of 82.4 mm fills about half of it. At the far end of the window, 800 mm, the rate drops to (1440 × 8) ÷ (5.6 × 800) ≈ 2.6 px per mm and the square still covers 8.24 × 2.6 ≈ 21 px. A practical floor for corner detection is about 5 px per square, comfortable at 10 px.
Decision: between roughly 21 px and 68 px per square across the whole window, the scaled board is detectable everywhere the arm can reach — no second scaling pass, no lighting escalation. If your own sensor lands under about 10 px per square, shorten the working distance before touching the board size again.
The fix, step by step
- Open the controller log on the pendant, run the wizard once, and note the pose number whose inverse-kinematics request returns false.
- Rerun from two different start positions. The same number failing again confirms reach; a moving failure point sends you back to the table above.
- Print the full-size grid at 58.8% — laser printer, driver at 100%, "fit to page" off.
- Measure with calipers: five squares in one line must span 5 × 8.24 mm = 41.2 mm, checked across and down. More than about 0.2 mm off and you reprint with a corrected scale, because that 0.5% square-size error flows almost one-to-one into the calibrated translation.
- Mount the board flat and rigid, refresh the wizard's snapshot pose with the board well inside its standoff window, and press Calibrate. The capture discipline in robot mounted camera calibration: capture and solve applies unchanged.
- Confirm 18 of 18 pose captures in the log, then teach a part and check that the reported pick point is where the gripper physically lands before you trust automated picks.
Checks before you rerun
- Rigid board mount. The board must not sag or shift partway through the run; tape on a carton gives up halfway. A column from the camera stands range or the accessories finder holds it still, and the same hardware carries the camera when the set-up is inverted.
- Standoff inside the window. The near half of the camera-to-board range is where the scaled grid's poses live; the snapshot pose decides it.
- Matte finish, no reflections. Judge glare from the camera's own preview, not by eye, and watch the board through the last poses for the arm's shadow.
- A separate intrinsic target is sized differently. It must fill the image, per checkerboard calibration target size: how many squares and how big; machined targets take printer scaling out of that error budget at close range.
- Fine alignment under a fixed camera. A 60 mm manual XY and rotation stage centers a target without touching the camera.
- No wizard at all? If the camera runs your own routine — an 8.8 MP C-mount smart camera with a global shutter, for example — you choose the poses yourself. Smart camera hand-eye calibration when no wizard exists covers the method, and eye in hand calibration pose selection lists the robot positions that stay inside the envelope.
The hardware half of this list is short: a rigid column for the board, and a machined target when the print is the weak link.
- Type
- Camera stands
- Type
- Calibration targets
When the usual answer is wrong
- Chasing start positions. Four different starts in the tracked case changed only how many images were captured before pose 11 failed. The board generates the pose set; the start does not.
- Editing the board's colors. The detector reads grid geometry, and the colored circles on the artwork carry no meaning for it. Contrast, matte finish and flatness matter; the palette does not.
- Adding light when the log says false. Light fixes detection faults. When the inverse-kinematics answer is false, brighter light only gives you a sharper photo of an arm that cannot get there.
- Reprinting at a friendlier round scale. A print that measures to spec is correct; a print that measures something else is wrong, however tidy the number.
Quick checklist
- Log read, failed pose number recorded, same number from a second start position?
- Grid matched to the arm: 8.24 mm squares for a compact arm, full-size artwork for a longer reach?
- Printed at 58.8%, driver at 100%, "fit to page" off?
- Calipers: five squares span 41.2 mm, across and down?
- Board rigid, flat and matte, 20-80 cm from the camera, no glare and no shadow?
- 18 of 18 poses captured, then one physical pick checked against the reported point?
Need a target machined to size or a stand to hold the board? Ask us on the contact page or start from the stand step of the configurator.
Questions
Why does the same calibration pose fail every single run?
The wizard derives the pose set from the board it detects, so one board and one snapshot pose always generate the same 18 poses. Your start position only changes the path the arm takes to reach them. A failure that repeats at one pose number means that pose is unreachable: the inverse-kinematics solver returns false every time. Failure points that move between runs mean the board is being lost during capture instead.
Does moving the robot's start position closer fix it?
No. The reachability of the computed poses does not follow the start position: with the board unchanged, the pose that failed stays outside the envelope wherever the arm begins. In the case tracked here, several starting locations changed only how many images were captured before the abort. Change the board size, or the arm, not the start.
How should an 8.24 mm square be checked with calipers?
Measure five squares in one continuous span: 5 × 8.24 mm = 41.2 mm, checked across and down. Calipers resolve a 41.2 mm span far better than a single small square. Hold the reading within about 0.2 mm; that keeps the square-size error near 0.5%, which is roughly the error that would land in the calibrated translation.
Does the color of the circles on the calibration grid matter?
No. The detector reads grid geometry — corner spacing and layout — not color, and the colored fiducial circles on this artwork carry no meaning for it. What matters is contrast, a matte finish against specular hotspots, and flatness. Time spent recoloring the board belongs in the log and the calipers instead.
The log shows inverse kinematics false on a different pose each run. Same fix?
No. A moving failure point points at detection, not reach: focus drift, glare, or a shadow crossing the board late in the sequence. Inspect the last captured image in the wizard, matte the print, remove reflections and the arm's own shadow, and hold the board 20-80 cm from the camera. Then rerun and read the log again.