Diagnose an eye in hand pick point offset

A detected part can look correctly centered in the image while its calculated robot pick lies near the wrist or another unexpected location. For an eye-in-hand setup, that usually calls for tracing the coordinate chain before adjusting the taught position. This diagnostic sequence helps you distinguish a missing calibration transform, the wrong capture pose, image-coordinate errors and tool-frame mistakes. Start with recorded data and stationary observations rather than trial-and-error robot moves.

Freeze the evidence before changing offsets

Keep the robot stationary during initial checks, following the machine's commissioning and safeguarding procedures. Inspect calculated coordinates without commanding a pick. An apparently reasonable position can still contain an incorrect orientation or approach direction.

Save one diagnostic bundle containing:

• The original image and its capture timestamp.
• The camera identifier and active project name.
• The calibration file identifier, modification date and load status.
• The robot pose associated with that exposure, including its reference frame.
• The detected point before and after each coordinate conversion.
• The active tool, work frame and final commanded pose.

Record expected and calculated positions in the same frame. Subtracting a camera-frame coordinate from a base-frame coordinate produces a meaningless error vector.

If another project works with the same hardware, compare its configuration without overwriting either project. A working project provides useful evidence, but its calibration is transferable only when the camera, mounting geometry and frame definitions match.

Check the active calibration before recalibrating

A project-specific failure makes calibration selection and loading a high-value first check. It does not prove that the calibration measurements themselves are bad.

Verify the configuration in this order:

1. Camera identity: confirm the capture step uses the physical camera associated with the calibration.
2. Mounting relationship: confirm the camera has not moved relative to the flange or tool frame used during calibration.
3. Calibration mode: check that the project expects an eye-in-hand relationship, not a fixed-camera relationship.
4. Transform endpoints: identify exactly which frames the stored matrix connects.
5. Runtime selection: inspect the file actually loaded by the running process, not just the file selected in an editor.
6. Load status: check startup messages and diagnostics for missing files, stale cached data or rejected settings.

Relative file paths deserve attention when projects are copied. A valid-looking path can resolve differently under another working directory or runtime account.

If the software documents a reload-on-capture or reload-on-start option, check its intended behavior. Reloading can address stale cached data, but it cannot correct the wrong file, incompatible frame definitions or changed mounting geometry. Repeat the stationary test after a full application restart.

If mounting geometry has changed, follow the vision calibration recovery checks. When fresh calibration measurements are necessary, use deliberate eye-in-hand calibration pose selection rather than collecting nearly identical wrist poses.

Trace the complete camera-to-base transform

Choose one transform convention and write it down. Here, a transform named B_T_C maps coordinates expressed in camera frame C into robot base frame B. Use column vectors and apply the rightmost transform first.

For a flange-referenced eye-in-hand calibration, the position chain is:

A missing camera-to-flange transform effectively treats the camera frame as the flange frame. That can make a detection appear centered on the robot's own axes. However, an identity fallback, zero-valued detection or missing target frame can create similar symptoms. Check intermediate values before choosing a cause.

For every matrix, verify:

• The source and destination frames.
• Whether translation is expressed in meters or millimeters.
• Whether angle fields use degrees or radians.
• The documented orientation convention and axis order.
• Whether the robot pose describes the flange or active tool center point.

If calibration is relative to a tool frame, use that frame consistently or convert it explicitly. Do not combine a flange pose with a camera-to-tool transform as though they share the same endpoint.

A rigid transform must be inverted as a complete rotation-and-translation relationship. Negating its translation alone is generally incorrect. The coordinate-transform guide for camera rotation explains why rotation changes how translations are expressed.

If the detection never produces a valid camera-frame point, resolve the missing target-frame condition before investigating downstream robot coordinates.

Use a worked example to recognize the error

Consider a simplified calculation with aligned axes and identity rotations. This example illustrates arithmetic only; it is not a robot motion target.

Suppose the recorded values are:

• Flange origin in the base frame: (400, 100, 200) mm.
• Camera origin in the flange frame: (60, 0, 40) mm.
• Detected point in the camera frame: (20, -10, 300) mm.

With aligned axes, translations add directly:

The omitted mounting translation explains this particular error exactly. With nonidentity rotations, simple addition no longer works; use the full transform chain.

Next, compare observations from different approved wrist poses while the target remains fixed. Evaluate the calculated base-frame point without executing a pick. A correct chain should place the stationary target at the same base-frame location within your measurement tolerance.

These patterns rank hypotheses; none identifies a cause by itself.

Pair each image with the pose at exposure

An eye-in-hand image belongs to the camera pose at exposure, not the pose when detection finishes. Using the latest robot pose can introduce an error even when calibration is correct.

Check whether acquisition is live, software-triggered or replaying a saved image. A recorded image paired with the current wrist pose is not a valid capture pair. Keep saved images and their original pose records together.

During testing, stop at the observation pose, allow motion to settle and capture using the documented acquisition sequence. Use manual exposure settings for repeatable comparisons where appropriate. A live preview is not proof that the processing step receives the latest frame.

For moving acquisition, measure the timing relationship rather than assuming timestamps share a clock. Include exposure timing, trigger latency and any robot pose interpolation. The timestamp-based frame matching guide covers separating image identity from processing completion.

A hypothetical timing error shows the scale. At an angular speed of 20 degrees per second, a 50 ms mismatch corresponds to:

This estimate considers rotation only. Robot translation, target motion and exposure blur can add further error. If the stationary test passes but the moving test fails, investigate timing before recalibrating.

Check image coordinates before robot coordinates

A correct rigid transform cannot repair a point measured in the wrong image coordinate system. Trace whether detection coordinates refer to the original image, a cropped region, a resized image or a rectified view.

For an unscaled crop, restore the crop origin before applying an original-image mapping:

For resized or warped images, apply the corresponding inverse coordinate mapping. Image rotation, padding and rectification require more than adding a crop offset. Keep calibration intrinsics consistent with the image geometry, or use a documented mapping back to the calibrated image.

A sampling estimate helps distinguish pixel errors from large transform errors. Assume a hypothetical planar view spans 160 mm across 1600 pixels:

This is a local planar estimate, not a substitute for perspective calibration. Use the resolution calculator when checking the sampling budget:

Also confirm how depth is assigned. A two-dimensional image point does not determine a unique three-dimensional pick without a known plane, depth measurement or another geometric constraint.

Separate the detected point from the gripping point

The visual feature and the required tool pose may not share the same origin. A clip's visible center, fixture datum and gripping location can all differ.

Inspect the point before grasp offsets, approach offsets and tool compensation are applied. Then enable each stage separately in an offline calculation. Check whether an offset belongs in the object frame, camera frame, tool frame or base frame.

An object-relative offset must rotate with the object. Adding it directly along a fixed base axis produces an error that changes with part orientation.

Verify the active tool center point independently using the robot manufacturer's documented procedure. Also check the work frame selected by the receiving program. If coordinates are already expressed in the robot base frame, interpreting them as work-frame coordinates applies another transformation.

A low image reprojection error does not validate every link in this chain. Use the independent checks described in camera extrinsic calibration accuracy beyond reprojection error.

Inspect hardware only after the data path is understood

Check mounting security, focus and cable strain without changing the calibrated camera relationship. Image blur or physical movement can invalidate measurements, but neither explains every project-specific offset.

Do not replace the camera simply because one project fails. If hardware selection is genuinely required, compare area scan cameras through the camera finder. Match shutter behavior, sampling and interface requirements to the acquisition task.

The MV-CU020-90GC is a global-shutter color camera. Its catalog characteristics provide a starting point for hardware comparison, not a remedy for a coordinate-chain error.

A separate bench fixture can help investigate image-coordinate behavior without moving the robot-mounted camera. Browse camera stands and positioning stages using the accessory finder. Check travel, load, mounting and repeatability requirements before selecting a fixture.

The LS-MNT-0220 is a manual XY and rotation stage. Consider it only for a suitable separate bench setup; its catalog description does not establish measurement accuracy or suitability for robot mounting.

Do not loosen the calibrated camera to make a bench experiment easier. Any change to its robot mounting relationship requires reassessment of the extrinsic calibration.

Verify the correction and prevent recurrence

Change one cause at a time, then repeat the original recorded calculation and a fresh stationary capture. Define an acceptance tolerance from the actual gripping task before evaluating the correction.

Use this final checklist:

• Identity: the runtime camera and calibration identifiers match the intended setup.
• Frames: every transform has documented endpoints, units and direction.
• Timing: each image is paired with its capture pose.
• Geometry: crop, resize, depth and grasp offsets are accounted for.
• Consistency: a fixed target remains fixed in base coordinates across approved observation poses.
• Persistence: the correction survives closing and restarting the application.
• Safety: calculated positions and orientations pass offline checks before safeguarded motion verification.

Archive the accepted configuration with its camera identity, mounting reference, tool definition and diagnostic capture bundle. Keep a known-good dataset for checking future project changes.

If you need help separating a camera-selection issue from a transform problem, send the frame definitions, anonymized intermediate coordinates and acquisition requirements through application support. Resolve the stationary coordinate chain before progressing to motion.