Robotic At-Line Metrology: When to Add a Camera-Based Optical Station

Evidence disclosure: this page is an independent engineering guide informed by public company information from Eleven Dynamics AG. It is not a customer case, an installed-base claim, an endorsement or a verified compatibility statement. No POMEAS camera, telecentric lens or measurement head is identified as installed on the referenced platform, and integration or purchasing status is UNKNOWN.

"At-line metrology" is a phrase that hides a decision. A robot cell that carries a part past a camera can produce an image, and the image can be processed into a number, but the number is only a measurement if the whole chain from part datum to reported value has been characterised. The question worth asking before adding a camera station is not whether the robot can reach the inspection position. It is whether the combined uncertainty of a camera-based station can be kept small enough, relative to the tolerance, to be worth the cycle time it consumes.

This page answers that question in the order it should be asked: uncertainty first, geometry second, cycle time third, and interface ownership last. It also states where a camera station is the wrong instrument, because a robotic cell is an expensive place to discover that the part presentation moves more than the tolerance allows.

A robot is motion, not metrology

A six-axis arm is specified by repeatability, usually in the tens of micrometres, and by absolute accuracy, usually an order of magnitude worse. Neither number is a measurement uncertainty, because neither describes the geometry of the feature being measured. When a camera is mounted on the arm, the arm's positioning error enters the chain twice: once as a change of viewing angle relative to the part, and once as a change of working distance. If the part is instead held by the robot and presented to a fixed camera, the same errors appear as a change of pose in the field of view.

That is why a camera station on a robot has two possible architectures and they are not equivalent. A fixed camera with a robot-presented part can use a rigidly mounted optical chain, a fixed calibration and a stable illumination geometry, which minimises the optical uncertainty but requires the robot to present the part consistently. A camera on the arm can reach features that a fixed station cannot see, but its calibration has to account for pose, and every pose becomes a calibration state. The first architecture is easier to qualify; the second is easier to justify on reach.

What an open platform actually integrates

Eleven Dynamics publicly describes EDAS as an open automation platform that connects machines, sensors and production data for inline and at-line metrology. The word "open" describes the software's ability to exchange data with devices; it does not establish native compatibility with any particular optical controller, trigger scheme or fieldbus, and it certainly does not establish that a POMEAS measurement head has been integrated. Those questions have to be answered by both vendors in writing, with a protocol and a synchronisation diagram, before a quotation is meaningful.

Four questions that decide whether a camera station belongs

First, is the part mechanically located or does it arrive in an arbitrary pose? A camera measures what it can see from one direction, so features on a hidden face need either a second view or a re-orientation, and the latter costs cycle time. Second, what is the tolerance and what fraction of it may measurement consume? A camera station that consumes most of a tolerance has removed the process margin it was bought to protect. Third, what cycle time is available per part, including settle, exposure and processing? Fourth, who owns the coordinate frame and the reject decision when the platform, the camera and the PLC all have an opinion?

A negative answer to the first question is not fatal, but it changes the design: a locating nest or a kinematic stop is usually cheaper than a vision algorithm that copes with arbitrary pose, and far easier to validate. A negative answer to the fourth question is fatal in practice, because two systems that both claim to decide pass or fail will eventually disagree on a real part.

The uncertainty budget, done with numbers

Take a hypothetical tolerance of plus or minus 0.10 mm on a planar dimension. Suppose the camera calibration contributes a standard uncertainty of 0.015 mm, the fixturing and part location 0.025 mm, and thermal effects over the working range 0.010 mm, treated as independent. The combined standard uncertainty is the root sum square.

Against a plus or minus 0.10 mm tolerance, an expanded uncertainty of 0.0616 mm is 31 percent of the tolerance width, or 62 percent of one-sided tolerance. The conventional rule of thumb is that measurement uncertainty should consume no more than about a tenth of the tolerance, so this budget does not pass, and the arithmetic says why: fixturing alone contributes 0.025 mm as a standard uncertainty, which is a quarter of the tolerance before anything else is added. The answer is not a better lens. It is a locating nest, a repeatable datum, or a looser tolerance on that particular dimension.

Note what the budget excludes: robot pose, algorithm edge detection, illumination variation and part-to-part form error. Those belong in the same table when they are measured, and their absence is the reason the numbers above are illustrative rather than certifying. A budget that only lists the optical terms is a budget that has already assumed away the difficult half.

Geometry: what a telecentric field removes from the budget

One term the optics can genuinely remove is the scale change caused by working distance. A telecentric lens holds magnification constant over its depth of field, which is why a part presented at a slightly different height measures the same size. The published POMEAS values make the trade-off explicit: a 1.0X bi-telecentric lens with an 18.4 mm object field holds 0.75 mm of depth of field at 110 mm working distance with telecentricity below 0.1 degree and distortion below 0.1 percent, while a larger field costs depth of field and a shorter working distance costs clearance.

That depth of field is the number that has to be compared with the robot's presentation error, not with its repeatability specification. If the part can arrive a millimetre higher or lower than taught, a 0.75 mm depth of field is not enough and the station needs either a mechanical stop or an autofocus mechanism, both of which change the cycle time. If the presentation is held by a nest, 0.75 mm is generous.

For measurements across a range of sizes rather than one fixed dimension, a motorised zoom is the practical compromise: a 0.68X-5.0X range at an 80 mm working distance covers several part families in one station. The condition, again, is that the zoom position is repeatable and the scale at each position is calibrated; a zoom whose magnification is known only to a few percent is a viewing aid. Where scale stability matters more than range, the 0.5X-1.0X telecentric macro zoom keeps distortion below 0.02 percent with a working distance between 75 and 106 mm.

Cycle time arithmetic

Consider an illustrative cycle: 2.0 s of robot motion to the inspection position, 0.3 s of settle time, an exposure in the tens of microseconds, and 0.15 s of image processing and result transmission. The station occupies 2.45 s per part, which is 1,469 parts per hour if the cell never waits. At a realistic 80 percent availability that becomes about 1,175 parts per hour, and if the line runs faster than that, the camera station becomes the constraint.

Two conclusions follow. Motion and settle dominate the budget, not imaging: the exposure is negligible and the processing is small, so buying a faster camera rarely buys cycle time, while shortening the travel or stiffening the fixture often does. And a settle time that is specified rather than measured is a guess: vibration after a rapid move decays over a period that depends on the structure, and the only reliable way to set the trigger delay is to image a reference target repeatedly at decreasing delays until the measured value stops moving.

Interface and ownership boundaries

Write down, before installation, who provides the trigger, what the trigger latency is, what the result protocol is, where the calibration constants live, and which system decides a reject. The failure mode to avoid is a station where the camera computes a value, the platform applies a correction, and the PLC applies a second one: three systems each behaving correctly and a reported dimension that belongs to none of them.

Two further items belong in the same document. The illumination has to be under the control of the same trigger as the camera, because a strobe that fires at the wrong time produces a plausible image with the wrong scale. And the measurement result has to carry the calibration revision and the zoom position, if any, so that a result can be traced back to the state of the station that produced it.

Pilot protocol

Run repeated measurements on a reference part at multiple positions in the working volume and across at least two shifts. Log the robot pose, the ambient temperature, the exposure, the calibration date, the operator and the measured feature, because a drift that cannot be attributed is a drift that cannot be corrected. Compare the results with a traceable reference instrument, and measure the complete station cycle time rather than the imaging time. If the same part measured at the same position on two shifts disagrees by more than the budget predicts, the budget is missing a term, and the missing term is usually thermal or mechanical rather than optical.

Frequently asked questions

Does robot repeatability equal measurement accuracy?

No. Repeatability describes how consistently the arm returns to a taught pose; accuracy describes how close that pose is to the intended one in the machine frame. Neither is a measurement uncertainty for a feature, because both omit the optical chain, the part datum and the algorithm. A repeatable robot presenting a part to a poorly calibrated camera station produces a repeatable wrong number, which is the worst outcome because it looks stable.

Can an HM measurement head be mounted on any robot?

Mechanical and interface feasibility are unverified for any specific robot. The published HM series covers measuring ranges from 15 mm to 120 mm with repeatability between 0.05 µm and 0.2 µm depending on model, and a head at that performance level needs a mount whose deflection and thermal behaviour are known, plus a communication interface the platform actually supports. Both questions have to be answered by the two vendors rather than assumed from a mounting pattern.

How much of the tolerance may a camera station use?

A common working rule is that measurement uncertainty should consume about a tenth of the tolerance. In the illustrative budget on this page the expanded uncertainty reaches 0.0616 mm against a plus or minus 0.10 mm tolerance, which is 31 percent of the tolerance width and well outside that rule. The dominant contributor there is fixturing rather than optics, which is typical: the cheapest improvement is usually a better datum, not a better lens.

Is a camera on the robot arm better than a fixed station?

They solve different problems. A camera on the arm reaches features a fixed station cannot see, at the cost of a calibration state per pose and a larger uncertainty budget. A fixed station with a robot-presented part keeps the optical chain rigid and the illumination geometry stable, which makes qualification far easier, but requires repeatable presentation. Choose on the measurement, not on the mechanical convenience.

Next step

Send the feature drawing with tolerances, the presentation method, the available cycle time and the current measurement method. The first reply is a budget, not a lens recommendation, because the budget decides whether a camera station is capable before the optics are chosen. For the chain around it, the inline dimensional measurement system guide covers how a fixed station is configured, the inline diameter error budget works through the same contributors on a cylindrical feature, and the comparison of gauge, confocal and vision covers the instrument classes that compete for this job. The note on PLC integration and sorting covers the interface boundary, and the analysis of telecentric drift when the workpiece moves covers what happens when the presentation error exceeds the optical tolerance. The HM telecentric measurement system and the telecentric lens range carry the current model data, and the contact page is where the drawing goes.