Robot repeatability is not process capability

A repeatability number answers a narrower question

A robot's pose-repeatability figure describes consistent return under a defined robot test. It does not by itself say how close the robot comes to a programmed target, how accurately it follows a path or whether the configured process produces an acceptable part or result.

The buyer's requirement should begin with the end output: the feature, position, path, assembly, inspection result or quality characteristic that must pass, its tolerance and reference, and the conditions in which it must be achieved. Robot repeatability is one evidence input to that requirement—not a substitute for it.

Pose and path measures are not substitutes

ISO 9283:1998 remains the current published international standard for industrial-robot performance criteria and related test methods, although ISO's page shows that the edition is under review in 2026. Its public catalogue establishes the standards scope; the full paid text was not reviewed for this article.

ABB's current IRB 6760 specification separately reports pose accuracy, pose repeatability, path accuracy, path repeatability and stabilization time. ABB claims pose repeatability of 0.06 mm and 0.08 mm for its two documented variants while reporting path accuracy of 1.4 mm and 1.1 mm. Those different figures are a manufacturer-specific illustration of why the measures cannot be exchanged.

• Pose repeatability concerns repeated arrival around a position; it is not automatically closeness to the programmed position.
• Pose accuracy concerns position relative to the programmed target; it is not automatically the result along a continuous process path.
• Path accuracy and path repeatability concern motion along a path and may matter more for cutting, dispensing, welding, scanning or finishing.
• The finished output adds the tool, fixture, object, sensing, calibration, process and measurement system to the robot's motion.

Start with the customer's output tolerance

This makes the requirement comparable across proposed solutions. It also prevents a provider from answering an end-process question with the most favourable motion figure on a data sheet.

• Name the characteristic that will be accepted: for example placed-part pose, hole position, assembled gap, bead or cut path, surface result or inspection reading.
• State the nominal value, tolerance, unit, datum or coordinate frame, measurement location and consequence of a failed result.
• Identify whether the task depends on a point, a continuous path, stabilization time, force or process behaviour, or several of these together.
• Record part presentation, object variation, fixtures, calibration, vision correction and every material operating condition instead of assuming an ideal input.

Keep the complete error budget visible

Separate the robot's motion from base and mounting deflection, tool-centre-point calibration, end-effector compliance, fixture and workpiece variation, sensing and coordinate transforms, process forces, thermal effects, wear and measurement uncertainty. Mark an unknown contributor as unknown rather than assigning the entire tolerance to the robot figure.

Attach each input to the exact load state, speed, path, workspace region, approach direction, environment and software mode. ABB says its own ISO 9283 results can vary with position in the working range, velocity, arm configuration, approach direction, load direction and gearbox backlash. That is ABB manufacturer guidance for the named product, not an independently verified rule or result for another application.

Ask for evidence with its conditions attached

Universal Robots' current UR20 technical page, for example, lists ±0.1 mm pose repeatability per ISO 9283. That is a manufacturer rating for the documented UR20, not evidence that a configured gripper, fixture, vision system and process will hold an end-output tolerance of ±0.1 mm.

• Name the exact robot variant, controller, software and performance mode proposed.
• For every figure, state the performance criterion, standard edition, load, offset, speed, test plane, workspace and other material test conditions.
• Keep rated, typical, sample-average, simulated, trial-observed and production-observed values as different claim types.
• Record the configuration or operating changes that can alter the result and the recalibration or retesting they require.
• Do not subtract a brochure repeatability figure from the process tolerance and call the remainder a verified engineering margin; the measures and conditions may differ.

Test the configured process, not the brochure

Freeze the robot, controller, software, tool, fixture, sensor, program, workpiece range and measurement method before an acceptance run. Use representative positions, paths, speeds, loads, approach directions, part variation, shifts and environmental conditions, and preserve raw results, rejected outputs, interventions, measurement uncertainty and the approval decision.

ISO 22514-2:2026 addresses statistical procedures for capability or performance using measured continuous quality characteristics. ISO 22514-3:2020 addresses short-term studies on consecutively produced discrete parts and carries important limits: its public abstract says it is not suitable below 30 observations or where tool-wear patterns or autocorrelation are present. These are process-study scopes, not universal robot acceptance rules. A qualified quality or process owner should choose a method that fits the output and evidence available.

Can the measurement support the tolerance decision?

A trial can produce a tidy table of readings without establishing that its inspection method is suitable for the output characteristic. Name the drawing revision, units, datum, measurement location and limits first. Then record the instrument or vision configuration, reference and calibration chain, resolution, setup, operators, representative part range and uncertainty assessment. A bench gauge study is not automatically representative of the way the production result is measured.

ISO 22514-7:2026, published on 15 September 2026, addresses validation of a measuring system and measurement process for a specific task. Its public abstract describes a simplified uncertainty-based approach primarily for simple one-dimensional measurements. That scope does not make it a universal test for complex geometric results, nor does validating a measurement process establish production capability. ISO 10012:2026 concerns management of measurement systems more broadly; neither public catalogue page verifies a particular cell or gauge. The full paid standards were not reviewed for this guide.

• Ask the quality or metrology owner whether the selected method and uncertainty assessment fit this characteristic and the actual working conditions. NIST distinguishes repeatability, reproducibility, stability, bias and resolution; a short study can miss change over time.
• Agree a decision rule before interpreting readings near a limit. Record whether the result is accepted, rejected or indeterminate under that rule, who owns the risk and what additional measurement resolves an indeterminate case. ISO 14253-1:2017 addresses uncertainty-aware conformity decisions for geometrical-product-specification characteristics within its scope; it does not supply a universal rule for every robot output.
• Keep the raw measurements, exclusions, rework, interventions and configuration revision alongside the decision. If measurement suitability or the limit rule is unknown, mark the tolerance claim unverified; do not quietly count a borderline result as a pass or assert that an unknown measurement proves nonconformance.

Reopen the comparison after a measurement change — 10 October method note

If the camera algorithm, gauge setup or datum changes between trials, a difference in the readings may reflect the measurement change as well as a change in the process. Preserve the earlier result with its conditions. Do not silently merge the two datasets or describe a lower spread as a process improvement before the measurement question is resolved.

NIST's public gauge-study design guidance identifies artifacts, operators, gauges, parameter levels and configurations as study choices. Its stability guidance illustrates examining check-standard results across runs. These are reasons to examine the measurement conditions and time evidence; the pages do not establish equivalence between two robot-cell trials or prescribe a universal retest threshold.

ISO 22514-7:2026's public abstract describes validation for a specific measurement task, primarily for simple one-dimensional processes using a simplified uncertainty approach. A changed method needs a task-specific assessment by the qualified owner. This guide has not reviewed the full paid standard and does not derive a sampling plan or acceptance criterion from its catalogue entry.

RobotAtom's proposed change record pairs the earlier and current characteristic, drawing revision, units, datum and limits; method and version; calibration reference, uncertainty assessment and decision rule; working conditions; and raw-result locations and dates. This is a coordination checklist, not a standardized equivalence study. A missing reference remains open even if the operator believes the setup was unchanged.

Consider a hypothetical inspection cell whose vision software changes while the part drawing stays the same. Keep both software versions, measurement locations, raw images or readings, exclusions and trial dates. Ask the quality or metrology owner which earlier conclusions remain usable, whether a bounded comparison is justified and what additional evidence is needed. The worksheet cannot answer those technical questions from matching names.

Record the reviewer's name, date, evidence basis, limits and disposition. Results may need to remain separate, or a reviewer may allow a narrowly defined comparison. A completed form can still withhold that comparison. No typed decision establishes measurement suitability, process capability, product conformity or permission to operate a robot.

Use the paired change-comparison worksheet alongside the original tolerance record. It highlights missing pairs and different typed entries and downloads the stated restrictions. Matching text is not proof of unchanged conditions; retain both source datasets and the reason for any later comparison.

Keep short-term and lifecycle evidence separate

A short acceptance study cannot automatically establish performance across tool wear, thermal drift, maintenance, changeovers, recalibration loss, material variation or every future production condition. Define which longer-term observations matter and what threshold triggers correction or renewed testing.

Reopen the affected evidence after a material change to the robot or controller, software mode, tool, fixture, payload, tool centre point, calibration, sensor, program, speed, path, workpiece, process consumable, mounting or environment. Preserve the earlier result as version-specific history rather than carrying its pass into a changed configuration.

Keep manufacturer claims and verification distinct

ABB and Universal Robots are authoritative for what they publish about their named products and stated conditions. Their ratings are manufacturer-provided evidence, not independent verification of a buyer's configured process. ISO is authoritative for the public scope and status of its standards, but the cited ISO pages do not verify a named robot or deployment.

RobotAtom records the required output, exact source, claim type, test conditions, configured-process evidence and unknowns. It does not certify robot accuracy, process capability, safety or compliance, and a performance test does not replace applicable safety, metrology, quality or sector review.

Make the evidence gap reviewable

The process owner and quality owner should be able to trace one claimed output back to the frozen cell, the actual measurement method and the agreed limit decision. Use the measurement-evidence record to identify what is documented, what still needs review and which change would reopen the claim. It is a scoping aid, not a capability index, conformity certificate or automated pass/fail verdict.