Define a new work plane after one full wrist camera calibration

The wizard refuses the working location — not enough reachable poses, board too far from the robot — and the natural question follows: can one full calibration, done somewhere the arm reaches comfortably, serve the real station? It can, and the sequence is the intended one, not a workaround. This guide gives the order of operations, the working-distance and pixel-size numbers your snapshot position must meet, the two-line fix when the robot drives the wrong way after detection, and the checks that prove the plane before the first real pick. It is written for anyone teaching locate-and-pick with a wrist camera from a teach pendant.

Can you run the full calibration at an easy pose and add the work plane later?

Yes — that is the intended sequence. The full calibration measures the fixed transform between camera and tool flange; that number travels with the wrist everywhere, so you can measure it anywhere the arm reaches comfortably and tie it to the working location afterwards.

An eye-in-hand camera is calibrated against the flange, not against the workspace. Every pick chains three transforms:

The middle term is the only thing the full calibration produces. The controller supplies the first term from its kinematics every cycle, and the detection supplies the third. A work plane never touches the middle term — it only states which physical plane the detections land on. That is why a new plane costs one board image, and why the calibration location is yours to choose.

Where does the board go when the working location is out of reach?

On a rigid surface close to the robot, at the camera's set working distance, where the wrist can orbit the board with tilt changes in every direction. Reach, not realism, is the constraint — the board never has to be at the working location.

Both refusal messages mean the same thing: the board sits outside the arm's dexterous zone. Fix the geometry. Bring the board inward until the wrist can circle it while the camera holds its working distance. If the grid is too large for that combination, use a smaller grid — calibration pose unreachable: scale the grid to robot reach works the arithmetic.

Pose variety beats pose count. Tilt the camera roughly 30 to 45 degrees off perpendicular between captures and rotate about the board normal; near-identical top-down shots add almost nothing to the solver. Eye in hand calibration pose selection: robot positions that work lists the orientations that make the solve converge in one pass.

Lock the board down — any shift between poses goes straight into the transform — with a camera stand sized in the stand configurator. The LS-MNT-0040 is a professional slide-table stand with a 700 mm height and a 400 x 450 mm base, enough table to clamp the board at the height the working distance demands.

How do you define the new work plane and snapshot position?

Move to the working location, lay the board flat on the actual part plane, capture one image, and store the pose you captured it from as the snapshot position. The full calibration must already exist and be saved before you start.

1. Finish the Full Calibration at the easy location, then save the program. The reference graphics refresh only after a save, even though the running code already uses the correct values.
2. Jog the wrist to the working location and frame the entire pick area plus about 20% margin — the field-of-view numbers below tell you whether it fits.
3. Lay the board flat on the plane the parts sit on, at part-top height, not on the fixture below them. For inclined or stacked parts, teach objects on an inclined work plane: calibrate at part-top height explains why the top surface is the plane you want.
4. Capture the single board image. The camera computes the plane equation from that one frame.
5. Store the pose you captured from as the Snapshot Position. The robot must return to it before every detection — camera locate snapshot position: why the robot must return first covers what breaks when it does not.
6. In the camLocate node, confirm the feature point it updates, then select that same feature point as the move feature in the pick move.
7. Save, reopen, and check the pendant graphics against the taught values.

Which camera do you mount when you build the wrist from parts?

A compact smart camera with its own light source, an M12-class lens and a field of view that frames your pick area at the working distance you can hold. Integrated lighting matters most — it removes a second fixture from the arm.

The full-calibration-then-work-plane sequence does not care whether the camera is a purpose-built wrist unit. Any camera whose detections the robot program can read works: run the same chain, teach the plane from one board image, take the part pose the same way. Filter the smart cameras by mount and resolution in the camera finder.

The closest match to an integrated wrist camera is one with its own multi-color light source, so the arm carries no separate illuminator:

Two more compact options for flange duty:

What working distance and pixel size does the new plane need?

Put the plane at the working distance the camera was built for, within about ±20 mm of it. Farther away, the field of view grows, the pixel gets coarser and the fixed-focus lens runs out of depth of field — all three at once.

As a worked example, take a 1/2.4-inch-class sensor about 5.3 mm wide, an 8 mm lens, 3072 pixels across the sensor, and a 300 mm working distance:

• Field of view: 5.3 mm × 300 mm / 8 mm = 199 mm across. A 150 mm calibration board fills three quarters of it.
• Pixel scale: 199 mm / 3072 px = 0.065 mm per pixel, so 65 µm. A 3 mm feature spans 46 px — resolution is not your limit.
• Depth of field at f/4 with a two-pixel circle of confusion (pixel pitch 5.3 mm / 3072 px = 1.7 µm, so c = 3.5 µm): 2 × 300² × 4 × 0.0035 / 8² = 39 mm, call it ±20 mm of usable focus.
• Teach the plane at 350 mm instead and the field of view becomes 5.3 × 350 / 8 = 232 mm, the pixel scale 0.076 mm — 17% coarser — and the plane sits 30 mm outside the focus window, so the board dots go soft and the plane solve weakens.

Hold the snapshot position at the designed distance. A 199 mm field covers a 160 mm tray with 24% margin (160 mm × 1.2 = 192 mm, which fits); anything wider needs lower magnification or a second snapshot position. Run your own camera through the calculator:

Keep the snapshot exposure at the value the Full Calibration used, so the plane image and the part images share contrast; wrist camera resolution troubleshooting covers the cases where the dots read grey.

Why does the robot move the wrong way after detection?

Because the pick move runs in the wrong frame: the feature point that camLocate updates is not selected as the move feature. Select it and the direction corrects itself — no recalibration, no new work plane.

The detection writes the part pose into its feature point, and the move after it must reference that point. Programmed against the base or tool frame instead, the arm drives a mirrored or opposite path — the classic symptom, and the first thing to check. Open the pick move, set its feature to the camLocate point, and re-run at low speed.

If the direction is right but the approach reports unreachable, look at the tool z-axis next: a 180-degree flip there is the next most common cause of unreachable-pose errors on the approach, and wrist camera motion error troubleshooting walks the remaining causes.

Why is the full calibration point at the robot base instead of on the board?

Because it is the camera's internal reference origin, not a physical landmark. Its coordinates are arbitrary — often near the base — and it is only the origin the detection poses are expressed against; it does not need to sit on the board or the plane.

The point moves on its own during execution: after every object detection its position updates, which is why it appears to wander in the pendant graphics. That is normal.

When you want to see what the camera sees, create a feature point and assign it the variable object_location: the pendant then draws the part pose the camera last reported — the fastest diagnosis for a direction problem, and a clean source for offset moves. To pull detections out of the pendant program entirely, camera image access over RTDE shows the channel that carries them.

When must you redo the full calibration instead of just the plane?

Only when the camera physically moves on the flange: remounting, a loosened screw, a collision, a swapped lens or refocused optics, or a changed bracket. New parts, new planes, new lighting — none of those touch the transform.

Check it without a board: park at the snapshot position, detect one stationary part ten times, and read object_location. At 0.065 mm per pixel, a 1 mm drift is 15 pixels — plainly visible. Scatter beyond about 1 mm or 0.5 degrees at a fixed pose means the mounting has moved; retighten, then recalibrate. For a permanent check between changeovers, robot cell vision recalibration with one reference target builds the same test around one glued target.

More stations do not mean more calibrations. Each new plane is one board image at its own snapshot position, and where several planes share a field of view, one snapshot position for multiple planes: apply an offset avoids re-teaching each one.

What checks confirm the new plane before the first real pick?

Run detection on a stationary part from the taught snapshot position, read object_location ten times, then make one reduced-speed approach. If the scatter stays under about 1 mm and 0.5 degrees and the approach stops on the part, the plane is good.

• Program saved and reopened; the pendant graphics match the taught values — a stale view hides a half-taught plane.
• Calibration board removed or parked outside the detection region; production detects parts, not the board.
• The camLocate feature point selected as the move feature in every vision-referenced move.
• Ten stationary detections within about 1 mm and 0.5 degrees; more scatter means mounting drift, not a plane problem.
• First cycles at reduced speed with nobody inside the guarded space; hand over to full speed only after the approach has stopped on target repeatedly.

If reach still blocks the full calibration after moving the board and scaling the grid, send the arm's reach and the camera's working distance to our applications team and we will size the board, poses and standoff together. More vision systems guides cover the calibration and plane-teaching workflow end to end.