Vision pick error under rotation: diagnose position drift

A pick that works near the taught orientation but misses after rotation usually needs a coordinate-chain test before another calibration. The reported part position, the requested robot position and the physical tool position are different measurements. Comparing them reveals where the error enters.

This procedure is for planar picking, primarily with a fixed camera. You will rotate a part around a controlled datum, separate translation from tool rotation, then repeat measurements across the working area. If the position is correct but the gripper makes an unnecessary full turn, use the separate gripper rotation troubleshooting procedure.

1. Classify the error before changing settings

Start with stationary parts and a repeatable fixture. Stop the feeder using the machine's approved procedure so motion and stale results cannot obscure the geometry.

Use this test order to investigate reference origins, transform calculations, tool offsets and calibration-plane errors. It is a diagnostic sequence, not proof that the first cause is always most likely.

Keep both components of the error vector. A distance alone hides whether the error rotates, reverses direction or remains fixed in the workspace.

2. Run a controlled rotation test at one location

Choose an identifiable pick datum, such as a hole center or a defined point on a calibration fixture. Do not assume a bounding-box center represents a fixed material point. Its position can change as the visible outline changes.

1. Secure the part so the chosen datum remains stationary during rotation. Verify that location independently of the vision result being tested.
2. Capture the reference orientation and save the image, detected pose and calculated robot target.
3. Repeat at approximately 45°, 90°, 180° and 270°, where the part's geometry permits unambiguous orientation detection.
4. Take at least five acquisitions at each orientation without moving the part between acquisitions. Record the spread as well as the mean.
5. Return to the initial orientation and confirm that the result returns to its original position.
6. Inspect the calculated targets before authorizing any robot movement.

Record the reported feature origin, reported pick point, angle, active recipe, calibration identifier, tool frame and work frame. Include acquisition identifiers so each target can be matched to its image.

Separate repeatability from systematic drift. If repeated readings vary by 0.1 mm but their mean shifts by 1.5 mm after rotation, random detection noise does not explain the main error.

For symmetric parts, an apparent 180° reversal may be a valid alternate match. Use an asymmetric feature or a suitable test artifact before interpreting that result as a transform failure.

3. Test the vision origin and local pick offset

A detected feature origin does not have to coincide with the intended grasp point. However, any displacement between them must rotate with the part.

Display the model origin, orientation axis and calculated pick point together. The procedure for adding camera preview overlays helps make these separate quantities visible.

For a planar system, a useful coordinate-chain model is:

Here, p_V is the detected model origin in calibrated vision-plane coordinates. The vector d_P runs from that origin to the desired pick point, expressed in the part frame. The rotation R(θ) turns that vector with the part. The rotation R_BV and translation t_BV map vision-plane coordinates into the robot base frame.

This equation assumes compatible length units and a calibrated planar mapping. Do not substitute raw image pixels directly into a rigid transform expressed in millimeters.

If the part rotates about its intended pick point, the detected model origin may legitimately move. The calculated pick point should remain stationary. Conversely, holding the model origin fixed means an off-center grasp point should trace an arc.

Check whether the software already applies the local grasp offset. Adding it again in the robot program produces another rotation-dependent displacement.

4. Use the drift magnitude to estimate an offset

A misplaced local origin or incorrect local offset often produces an arc-shaped error. For one erroneous offset vector, the change from the reference orientation follows this idealized relationship:

The vector d represents the offset error at the reference orientation. The angle θ is rotation relative to that orientation. This model assumes a rigid planar transform and excludes detection noise, calibration distortion and tool errors.

For an offset magnitude of 2 mm:

• At 40°: displacement = 2 × 2 × sin(20°) ≈ 1.37 mm.
• At 90°: displacement = 4 × sin(45°) ≈ 2.83 mm.
• At 180°: displacement = 4 × sin(90°) = 4 mm.

Under this model, a 4 mm change between opposite orientations suggests an offset error near 2 mm. That estimate tells you where to inspect; it is not a correction value to enter without verification.

Plot the horizontal and vertical errors against angle. A smooth rotating vector supports an offset hypothesis. Abrupt jumps suggest alternate feature matches, recipe changes or angle discontinuities instead.

5. Audit the transform between vision and robot

If the reported pick datum remains stable but the calculated robot target moves, inspect the coordinate conversion before recalibrating the camera.

Check these items individually:

• Absolute pose versus relative offset: Determine whether the vision result is a complete location or a correction relative to a taught reference.
• Transform direction: Confirm whether each transform maps vision to robot coordinates or robot to vision coordinates.
• Multiplication order: Applying a translation before rotation generally differs from applying it afterward.
• Axis direction: Image vertical coordinates commonly increase downward, unlike many planar work-frame conventions.
• Angle units: Confirm degrees or radians at every interface.
• Offset ownership: Assign each translation and rotation correction to one stage, not both vision and robot logic.
• Reference consistency: Match the vision reference image to the intended robot reference pose and active recipe.

Use a known point and a known direction to verify the transform. Agreement at one point cannot establish correct axis directions, scale or rotation.

For changeover-dependent faults, review vision project ownership during changeover. A correct transform used with the wrong reference data still produces an incorrect target.

Do not discard measured translation simply because a rotation-only test appears better. That removes the ability to pick parts at arbitrary positions and can conceal the underlying frame mismatch.

6. Separate the TCP from the vision result

Next, distinguish a wrong requested pose from a tool that does not physically realize the requested pose.

First compare calculated positions while holding the requested tool orientation constant. If the targets already drift with part angle, tool rotation is not required to produce the fault. Continue investigating the vision result and transform chain.

Then use the manufacturer's documented TCP verification procedure. At an approved verification location, compare the physical tool point across permitted orientations while the controller maintains the same commanded TCP position.

An incorrect TCP can make the physical tip move even when the displayed target coordinates remain unchanged. A tilted tool axis can also couple orientation changes into lateral or height errors. A visually concentric movement from one viewing direction does not verify the full three-dimensional tool definition.

Check for mechanical movement after the mathematical tool definition is verified: loose fingers, compliant suction hardware and changing contact geometry can alter the effective grasp point.

For a robot-mounted camera, also log the robot pose at image acquisition. Camera-to-tool calibration and the acquisition pose become additional parts of the coordinate chain.

7. Repeat across the field and at the working height

Once the central rotation test passes, repeat it at the center and four separated positions across the usable field. Keep the same target height and angle convention.

An origin-offset error tends to retain a similar relationship to part angle across the field. Calibration errors more often change with image position, although several faults can coexist.

A conventional perspective camera calibrated on one plane does not automatically provide correct planar coordinates for another height. For an ideal camera looking perpendicular to a plane, a small height mismatch gives this approximate lateral error magnitude:

For example, at a radial distance of 150 mm, a height mismatch of 5 mm and a camera height of 500 mm:

This is a small-height approximation, not a correction formula for an oblique installation. It explains why a central test can pass while edge picks fail.

Review camera calibration for length measurement for coordinate-mapping checks, and calibration target flatness when the calibration surface may be distorted.

Calibrate or validate at the actual feature plane. If part height varies substantially, investigate a suitable three-dimensional measurement approach rather than dividing the field into regions to hide the error.

8. Check the architecture before replacing hardware

Replacement hardware is not the first remedy when a stationary, sharply imaged datum produces repeatable angle-dependent drift. Establish which stage changes the point before choosing parts.

For an on-device inspection architecture, review smart cameras and vision sensors. The MV-SC3013XM-12M-ATN-H is a catalog candidate; confirm access to the required pose results, reference settings and diagnostic overlays before selection.

For host-side image processing, compare area scan cameras through the camera finder. The MV-CU020-90GC is another catalog candidate. Camera selection alone does not establish robot-coordinate accuracy.

Where separate processing hardware is required, evaluate vision controllers. Review the MV-SVC2506-16G60-128G against the intended acquisition and software architecture rather than assuming application compatibility from its category.

Before ordering, confirm image access, pose conventions, calibration workflow, robot communication and any required software licensing. Ask specifically whether diagnostic coordinates are available before and after each transform.

9. Verify the correction and prevent recurrence

Change one cause at a time, then rerun the same saved images and physical tests. Recalibrating everything at once can remove the symptom without showing which setting caused it.

Use this acceptance checklist:

• Rotation sweep: The intended pick datum remains within the application's tolerance across all required orientations.
• Field coverage: The same criterion passes at separated locations throughout the usable working area.
• Height coverage: Representative part heights pass, or the recipe explicitly restricts the valid plane.
• Tool verification: The physical grasp point agrees with the commanded TCP under the required orientations.
• Restart and changeover: Correct calibration, reference, tool and work-frame data return after normal operating transitions.
• Traceability: Save representative images, calculated targets, residual errors and configuration identifiers.

Define tolerance from the available contact margin, not from what looks centered on a preview. For an 8 mm landing width and a 4 mm contact width, nominal lateral clearance is (8 − 4) / 2 = 2 mm per side. Part variation, tool compliance and other errors consume that allowance.

Do not call the correction complete after one successful pick. Repeat the angle and field tests, then verify the intended operating cycle under approved conditions.

For a new installation, start with the camera configuration workflow. For an existing cell, send the rotation-test results, frame definitions and representative images through application support so the next decision follows the measured failure.