Robot vision for locating part center points

A camera does not give a robot a usable pick position simply by finding the middle of an image blob. You need an agreed part reference, calibrated coordinates, predictable imaging, and a controlled handoff between detection and motion. This guide is for integrating a camera with an existing robot to locate parts. Start with the simplest arrangement that fits the task: a fixed camera viewing stationary parts on a known plane. Add tracking, additional views, or depth measurement only when the application requires them.

Define what the center point means

Decide which physical point the robot needs before selecting hardware. The center of a visible outline, a fitted hole, and a usable gripping surface can occupy different positions on the same part.

Common definitions include:

• Silhouette centroid: the area-weighted center of the segmented outline. Missing material, shadows, and partial occlusion can move it.
• Fitted geometric center: the center of a circle, rectangle, or other fitted shape. Its reliability depends on visible edges and the fitting method.
• Model reference point: a taught location attached to a recognized part pattern. This can represent a grasp point away from the visible center.

Write the output requirement as a coordinate contract. For example, request the grasp reference position in the robot base frame, with an orientation convention and a separate approach height.

A center point alone does not describe rotation. If the gripper must align with a long edge, also estimate the part angle. Symmetric parts may require an additional identifying feature to resolve orientation.

For the underlying planar-system decision, use the 2D vision pick-and-place selection guide. Keep this workflow focused on turning the chosen reference into a verified robot action.

Choose the simplest camera placement

A fixed overhead camera is usually the starting point for parts presented on a flat tray or stopped conveyor. It keeps the camera pose constant and separates image acquisition from robot positioning.

Place the camera where the gripper does not obscure the part during acquisition. Include the largest part, presentation variation, and usable tray edges in the field of view. Check that tall features do not hide the reference you intend to detect.

If parts move between imaging and picking, camera position becomes part of the tracking design. The conveyor camera-placement guide addresses that layout decision.

Ordinary vision detections are not safety-rated protective functions. Commission robot motion using the documented safeguarding and operating procedures for the complete machine.

Turn pick tolerance into image requirements

Start with the allowable error at the gripping feature, not a preferred camera resolution. Allocate that tolerance among image localization, calibration, robot positioning, tooling, and part presentation.

For a nearly perpendicular view of a flat plane, object-space sampling can be estimated from field width and image width:

These relationships describe sampling, not guaranteed measurement accuracy. Lens distortion, focus, edge contrast, and perspective still matter. Sampling can also vary across an oblique view.

Consider a hypothetical field measuring 240 mm by 180 mm, sampled by an image measuring 4000 by 3000 pixels:

That feature is sampled across several pixels, but its visibility still depends on contrast and optical sharpness. Do not treat the arithmetic as proof that the system will meet a 0.30 mm pick tolerance.

Use the resolution calculator to explore the trade-off between coverage and feature sampling:

For moving parts, estimate exposure blur separately:

Shorter exposure reduces this blur but requires sufficient light. A global shutter avoids rolling-shutter geometric skew; it does not eliminate blur during exposure.

Select hardware around the imaging conditions

Use area scan cameras for complete images of the pick area. The camera finder helps narrow interface, shutter, and sensor requirements after you establish coverage and timing.

The LS-CMR-0166 is a USB3 rolling-shutter camera to consider for a stationary-part evaluation. Confirm lens compatibility, acquisition controls, and integration support before selection. Do not assume it suits moving parts simply because image transfer is fast enough.

Use bar lights when directional illumination can separate the part edge or reference feature from its background. Evaluate the LS-BAR-1047 as a candidate, rather than assuming its physical coverage guarantees suitable contrast. The lighting finder provides alternatives when the initial geometry produces shadows or reflections.

Stable mounting matters because camera movement changes the calibration. Explore camera stands and the accessories finder when assembling a repeatable bench test.

The LS-MNT-0040 is a camera-stand candidate. Confirm travel, mounting interfaces, stiffness, and clearance for the selected camera and optics before ordering.

These parts form an evaluation shortlist, not a complete, prequalified robot-vision package. Lens selection, triggering, communications, and mechanical integration remain application-specific.

Calibrate the plane and coordinate frames

For a fixed camera and a flat working surface, establish a mapping from image coordinates to coordinates on the physical pick plane. Correct lens distortion using the supported calibration method before relying on a planar mapping.

A planar projective mapping can be written as:

Here, u and v are image coordinates, X and Y are coordinates on the calibrated plane, and s is a scale factor. The mapping does not independently measure height.

Use known reference locations distributed across the usable area. Avoid collecting every calibration point near the center. A mathematically fitted transform can behave poorly near edges that were not adequately represented.

Follow the software's documented procedure for obtaining corresponding image and physical coordinates. When robot positions provide those physical references, verify the active tool, work frame, units, and reference feature.

The four-known-point plane-pose explanation provides background on planar geometry. For the calibration arrangement, consult the robot camera calibration explanation. Neither replaces independent verification of the finished station.

Make lighting and detection repeatable

Lighting should stabilize the feature used to locate the center, not merely produce an attractive image. For silhouette-based detection, shadows can merge with the part and move the apparent centroid. For pattern-based detection, reflections can hide or create edges.

Begin with fixed exposure, gain, aperture, and light settings during evaluation. Record them with the calibration. If automatic controls are necessary, test their full permitted range rather than calibrating under one favorable setting.

Compare these lighting changes before adding more image-processing stages:

• Change the illumination angle to move reflections away from the reference feature.
• Change background contrast to separate the full part outline.
• Add illumination from another direction when one-sided shadows bias the detected center.
• Reduce ambient-light influence through suitable lighting geometry and exposure control.

The lighting selection guide explains how to choose illumination around the inspection feature. If glare dominates, the cross-polarization guide describes another technique to evaluate.

Test representative good parts, finish variations, allowable rotations, and difficult presentation positions. A detector that works only on the cleanest sample is not ready for robot integration.

Define the detection-to-motion handoff

Treat each detection as a structured record, not a pair of numbers sent whenever the camera finishes processing.

Include the following information where the receiving system supports it:

• A unique acquisition or target identifier.
• Position, units, and an explicitly named coordinate frame.
• Orientation and its convention when required.
• A valid or invalid state and the reason for rejection.
• Acquisition time or tracking reference when the scene can move.
• Recipe or calibration identity where configuration changes are possible.

Define what happens when no part is found, several parts overlap, confidence falls below the acceptance rule, or the robot cannot reach the target. Do not let an old valid coordinate remain available as though it describes a new image.

A simple handshake separates acquisition requested, image processed, target accepted, and target consumed. After a pick, confirm that the next detection belongs to a fresh acquisition or to an explicitly managed target list.

Use the robot-base pick-coordinate guide for applying established transforms. If processing completes asynchronously, the frame-matching guide explains why result identity must remain attached to its source image.

Validate across the complete working area

Reserve physical reference positions that were not used to fit the calibration. Measure errors at these independent locations, including edges and corners.

A practical initial test uses a 5 by 5 layout, giving 25 positions across the usable plane. At each position, repeat acquisition with the part stationary, then repeat placement and detection. These tests separate image repeatability from presentation variation.

Calculate position error against the independent reference:

Report maximum observed error as well as a typical value. An average can hide a corner where every pick fails. Also check orientation error when the gripper depends on angle.

For preliminary budgeting, independent, approximately zero-mean error contributions expressed as comparable RMS values can be combined by root-sum-square. Suppose localization, calibration, robot positioning, and tooling contribute 0.12 mm, 0.18 mm, 0.10 mm, and 0.15 mm:

This is not a worst-case bound. Systematic offsets, correlated errors, and occasional detection failures require separate treatment. Do not substitute published repeatability or a calibration fit residual for a measured contribution without checking what it represents.

Finally, validate actual gripping using the machine's approved commissioning procedure. Image-coordinate accuracy alone does not verify tool offsets, approach direction, clearance, or reliable engagement.

Recognize when a planar center is insufficient

The simplest workflow is wrong when the task violates its assumptions.

Variable height: A part reference above or below the calibration plane can shift its apparent position. Constrain height, calibrate the relevant plane, or use suitable spatial measurement.

Tilted or stacked parts: A visible center does not establish the surface normal, accessible grasp face, or collision-free approach. Additional sensing and planning may be necessary.

Occlusion: A partial outline can produce a plausible but incorrect centroid. Use supported model features, reject incomplete views, or change presentation.

Symmetry: A circular outline cannot identify an arbitrary angular reference. Find an asymmetric feature or use a grasp that does not require that angle.

Changed geometry: Moving the camera, changing focus, replacing optics, or altering the mounting can invalidate the established relationship. The calibration recovery guide explains how to assess changes before resuming production.

Check the integration before ordering

Use this checklist to turn the concept into a testable specification:

• Part reference defined: State the exact point and orientation the gripper needs, including allowable ambiguity.
• Presentation bounded: Document height, tilt, overlap, and motion limits so the sensing method matches reality.
• Coverage calculated: Include the largest part and placement variation without sacrificing necessary feature sampling.
• Timing established: Separate exposure, transfer, processing, communication, and motion delays.
• Lighting demonstrated: Save representative images showing difficult finishes and orientations, not just ideal samples.
• Calibration ownership assigned: Define who maintains frames, tool offsets, calibration files, and change records.
• Validation method agreed: Specify independent reference positions, error limits, and required pick outcomes.
• Failure behavior defined: Reject uncertain, stale, unreachable, or duplicate targets through the normal control sequence.
• Safeguarding reviewed: Keep protective functions independent of ordinary camera detections unless the complete function is appropriately safety-rated.

Start the camera configuration workflow with your field of view, feature size, motion state, and interface requirements. For an integration review, send those requirements and representative part images through application support, together with the required pick tolerance and definition of the grasp reference.