Robot mounted camera calibration: capture and solve

A robot-mounted camera needs more than a sharp image and a calibrated lens. The controller must know how the camera coordinate frame relates to the robot frame carrying it, and every calibration image must match the correct robot pose. This workflow takes you from preparation through capture, solving, and independent validation. It assumes a rigidly mounted camera observing a stationary calibration target while the robot stops at each capture position.

What you need before capture

Prepare a camera with controllable acquisition, a rigid mount, a dimensionally known target, and access to timestamped robot poses. Your calibration software must estimate target poses and solve an eye-in-hand calibration using documented coordinate conventions.

Use area scan cameras with acquisition software that exposes the image settings and metadata you need. The camera finder helps compare connection and sensor options, but software compatibility and mechanical suitability still require confirmation.

The LS-CMR-0053 provides a GigE interface and a rolling shutter. For this workflow, acquire images only after motion and vibration have settled.

The LS-CMR-0166 provides a USB3 interface and a rolling shutter. Confirm acquisition support in your chosen software before selecting the camera.

These examples identify acquisition options, not qualified robot-mounted assemblies. Check mounting provisions, payload, cable routing, lens retention, and permissible operating conditions separately.

For offline optical checks, camera stands can hold the camera independently of the robot. The accessories finder helps select bench positioning hardware; a bench stand is not automatically a suitable robot bracket.

Plan calibration movements under the machine's approved setup procedure, with its protective measures maintained. Check the complete camera assembly and cable sweep throughout the intended capture workspace.

Define the frames and transform direction

1. Write down every coordinate frame before collecting data.

Use the following definitions consistently:

Each symbol represents a rigid homogeneous transform. With column vectors, the expected relationship is:

The same target point should therefore map to the same base coordinate from every capture.

If the robot reports an active tool frame rather than the flange, the solved transform relates the camera to that tool frame. Either use flange poses or document the tool transform and preserve it throughout capture and deployment.

Record translation units, angle units, rotation representation, axis directions, and multiplication order. A pose containing Euler angles cannot be interpreted correctly without its specified rotation convention. Likewise, a camera optical frame may differ from a general mounting frame.

A solver may request the inverse of a transform defined above. Convert explicitly at the software boundary rather than changing conventions halfway through the dataset.

Calibrate intrinsics and check target sampling

2. Establish the camera model before solving its relationship to the robot.

Calibrate focal parameters, principal point, and lens distortion at the image resolution and optical settings you will use. Lock focus and aperture afterward. Changing focus, resizing images, or changing the sensor crop can invalidate the relationship between image coordinates and the stored camera model.

For distorted images, use the corresponding distortion model during target pose estimation. For rectified images, use the rectified camera matrix and avoid applying distortion correction again.

Check the target's physical dimensions independently. The calibration target flatness guide explains why a dimensionally correct but bowed target can still produce inconsistent observations.

Estimate feature sampling before committing to the optical layout. Suppose the horizontal field of view is 240 mm, the smallest relevant target feature is 2 mm wide, and your planning goal is 20 pixels across that feature:

This is a sampling estimate, not a universal detection threshold. Confirm the detector's requirements and inspect images at the most oblique planned views, where projected features become narrower.

Leave room for the entire required target pattern and its identification features. The target field of view and coverage guide covers framing margins in more detail.

Use repeatable illumination and fixed exposure settings once a workable image has been established. If exposure is unnecessarily long, the front-lighting workflow for calibration targets explains how to improve capture conditions without relying on gain alone.

Capture matched observations at settled poses

3. Keep the target stationary and collect varied, matched image-pose records.

Place the target where the camera can observe it from a useful range of robot orientations. It must remain fixed relative to the robot base throughout the dataset. Moving it between captures breaks the stationary-target model.

At each capture position:

• Move using the approved calibration procedure and wait for settling.
• Acquire an image and associate it with the corresponding robot pose.
• Save the original image before processing or annotation.
• Detect the target and estimate its pose using the verified camera model.
• Check corner coverage, identification, sharpness, and pose plausibility.
• Accept or reject the complete image-pose pair together.

A robot pose read after image delivery is not necessarily the pose at exposure. For stopped acquisition, confirm that the robot remains stationary throughout both events. For moving acquisition, a validated synchronization and motion model is needed; that is a different capture workflow.

Store an image identifier, robot pose, camera settings, target definition, timestamps, and the association method. Do not subtract camera and controller timestamps unless their clock relationship is known.

An initial planning dataset might contain 24 valid captures: 18 for solving and 6 withheld for validation. This is an example allocation, not a guaranteed minimum. Geometric diversity matters more than repeatedly observing almost the same pose.

Include rotations about more than one nonparallel axis, along with useful changes in position. Pure translations or rotations around only one axis can leave parts of the full transform poorly constrained. The eye-in-hand pose selection guide explains how to improve that diversity.

If target detection intermittently disappears, follow the missing target-frame checks before collecting more observations.

Solve the hand-eye transform

4. Solve from the accepted training captures and preserve the complete result.

Some solvers accept absolute robot and target poses. Others form relative motions and solve the hand-eye equation. With the definitions above, two captures give:

These equations depend on the stated transform directions. A different software interface may define its relative motions differently, so match the documented input contract rather than copying matrix names alone.

Before solving, check that all records use the same target dimensions, camera model, robot reference frame, and translation units. Reject invalid rotation matrices and incomplete records. Inspect planar target solutions for flipped orientation or inconsistent target-frame assignment.

After solving, save the transform with explicit source and destination frame labels. Also save the training capture identifiers, software configuration, camera calibration, target definition, and residual report. An unlabeled matrix is not a deployable calibration record.

Inspect training residuals, but do not treat them as independent proof. A solver can fit internally consistent data that contains an incorrect target scale or robot frame convention. If estimated target distances look implausible, the fiducial distance troubleshooting guide helps separate scale and frame errors.

Validate with unused poses and independent geometry

5. Evaluate withheld captures before using the transform for guidance.

For each unused capture, compute the target pose in the robot base:

Because the target stayed fixed, these estimates should agree. Compare target-origin positions and orientation differences across views. Report both typical variation and the worst observed error, including where it occurred.

For example, suppose a target-origin estimate differs from an independent reference by 0.18 mm, 0.24 mm, and 0.12 mm along the base axes:

That result must be judged against the application's allocated calibration error, not an arbitrary universal tolerance. An inspection allowing 0.50 mm total error cannot necessarily spend all of it on calibration; robot positioning, feature localization, and tooling also contribute.

Agreement among views measures consistency, not complete absolute accuracy. Use independently surveyed target points or another traceable geometric reference to check base-frame accuracy. Do not derive the reference from the same observations being assessed.

Validate multiple target points and several working orientations. A small angular error can create a much larger position error away from the camera origin. Tests near one central point may conceal that effect.

The extrinsic calibration accuracy guide explains why low image reprojection error does not establish robot-coordinate accuracy.

Diagnose failed or misleading solutions

6. Isolate data problems before changing solver settings.

Use removal of observations sparingly. A visibly blurred image is a defensible rejection; excluding every pose that disagrees with the desired answer is not.

If a different solver produces a substantially different transform, compare the resulting predictions on the same withheld dataset. More agreement between solvers does not replace physical validation.

Release checklist and recalibration triggers

7. Release the calibration only after its operating conditions are documented.

• Confirm that deployed images use the calibrated resolution, crop, focus, and distortion treatment.
• Confirm that the robot pose source matches the frame used during solving.
• Verify transform direction and units in the consuming application.
• Retain raw captures, accepted-pair identifiers, and rejection reasons.
• Record withheld-pose results and independent reference measurements.
• Save a verification pose that can be repeated after maintenance.
• Define which mechanical or optical changes require verification or a new calibration.

A rigid, repeatable mounting change may support a controlled recovery procedure, while an unknown shift requires renewed measurement. The vision calibration recovery workflow helps decide what must be repeated.

If the camera, target, or robot cannot remain stable during stopped capture, do not force this workflow to fit. Resolve the mechanical or synchronization problem before trusting the transform.

Start the camera configuration workflow once the acquisition and mounting requirements are clear. For help checking camera compatibility and capture requirements, send those details through the applications support contact page.