Robot repeatability is not absolute accuracy: when should a cell be calibrated?
A robot may return tightly to a taught pose yet miss coordinates supplied by CAD, vision or another machine. Specify the spatial relationship the task actually depends on, then test repeatability, absolute positioning, frame registration and the final process separately.
The short answer: a fixed, taught process may depend mainly on repeatability because teaching absorbs a stable coordinate offset. Offline programming, vision guidance, robot replacement, bimanual handovers and mobile manipulation depend on external frames agreeing with physical reality, so absolute positioning and registration need their own acceptance evidence. Calibration can reduce systematic, modelled error; it cannot repair unstable mechanics, thermal variation, changing loads or poor repeatability.
“Robot accuracy” is at least four different questions
ISO 9283 defines performance criteria and related tests for manipulating industrial robots, including pose and path behavior. Those controlled robot measurements are useful inputs, but they are not a process-capability result for a complete cell.
| Evidence layer | Question it answers | Typical failure | Appropriate check |
|---|---|---|---|
| Pose repeatability | Do repeated commands form a tight cluster? | Scatter from looseness, backlash, temperature or changing load | Repeated measurements at relevant payloads, speeds, poses and approach directions |
| Absolute pose accuracy | Does the measured TCP reach the commanded coordinate? | Consistent bias from joint zero offsets, link parameters or deflection | Compare command and measured pose with an external reference |
| Frame registration | Do robot, tool, part, camera and peer-machine frames agree? | Incorrect TCP, base, work-object or hand-eye transform | Validate each transform and use independent check poses |
| Process capability | Does assembly, machining, inspection or handling meet its requirement? | Combined tooling, perception, contact, material and workflow error | Run the complete task on representative parts |
ISO 9787:2013 defines robot coordinate systems and motion nomenclature for alignment, testing and programming. Common names help teams communicate; they do not prove that base, flange, TCP, work-object and camera transforms have been measured correctly.
Choose the evidence from how points enter the process
| Programming and task mode | Evidence to prioritize | Why |
|---|---|---|
| Fixed fixture with every point taught in the cell | Repeatability and process output | Teaching absorbs stable bias, but not random spread or drift |
| CAD or offline-generated paths | Absolute pose/path accuracy and base registration | The real robot must reproduce coordinates created outside the cell |
| Vision-guided handling or inspection | Camera, hand-eye/workcell, TCP and task validation | Pixel-to-part-to-robot transforms all contribute to the landing error |
| Bimanual work or transfer between robots | Each arm's repeatability, relative registration and shared-part result | Small biases from two kinematic chains can meet at one contact task |
| Mobile manipulation, station reuse or robot replacement | Docking/localization, base, arm, tool and station registration | The system reconstructs a longer spatial chain after every move or change |
Calibrate systematic error only after stability is established
NIST's study on improving and quantifying robot accuracy in situ addresses the need for positional accuracy in simulation-assisted programming and high-tolerance assembly. By identifying joint zero offsets, the researchers improved average positioning accuracy roughly fourfold for the studied arm and measurement setup. That is evidence that in-situ kinematic calibration can work—not a transferable accuracy promise for every robot, payload or workspace.
NISTIR 8093 describes calibration as identifying a better relationship between joint readings and the end effector's actual workspace position, then applying the identified changes to positioning software. It also explains that coordination with external sensors, tools or other robots requires registration to an external coordinate frame. Kinematic calibration, TCP calibration and camera-to-cell registration should therefore remain separate line items.
A seven-stage calibration acceptance workflow
- Start from the process tolerance. Define the product result, critical poses and permitted compensation. Do not copy a datasheet repeatability value into the cell requirement.
- Freeze the test state. Record robot and software versions, tool, payload, speed, warm-up, temperature, mounting and safety configuration. Resolve evident mechanical or installation faults first.
- Measure repeatability before fitting a model. Exercise the task-relevant workspace, orientations and approach directions. Unstable dispersion points to mechanics, thermal state, load or measurement—not a richer compensation model.
- Establish a credible reference. Select external metrology that covers the required position and orientation range. Record its calibration state, setup, environment and uncertainty. ISO/TR 13309 addresses test-equipment principles and metrology methods used with ISO 9283.
- Separate the error chain. Check arm kinematics, base, TCP, work object, camera and mobile-platform localization individually. Adjust only the layer supported by evidence.
- Hold out validation poses. Identify calibration parameters with one set of points, then verify untrained workspace poses and paths. A fit that only works at calibration points is not cell-wide acceptance.
- Return to the process. Validate quality, cycle, clearance and recovery with representative parts. Define rechecks after relocation, collision, maintenance, tool change or a reference trend breach.
Treat calibration as controlled configuration, not a permanent label
NIST's work on industrial robot accuracy degradation monitoring discusses position-dependent errors caused by geometry as well as load, gravity, backlash, thermal environment and degradation. It also notes that an instrument that is not repeatable cannot be represented reliably by calibration or an error model. A plant can retain a small set of stable reference poses or parts for trend checks, using a breach to trigger investigation or revalidation rather than automatically declaring the robot uncalibrated.
Matrix Dimension perspective
The following is an engineering inference from the standards and NIST work: for an AI single/dual-arm control hub or wheeled humanoid project, the useful starting document is a transform chain from task coordinates to the business result—not a single “total accuracy” number. Mark who establishes each link, how it is measured, what invalidates it and how it is recovered. That map shows whether the next investment belongs in teaching, vision compensation, kinematic calibration, tooling or task design. Our embodied-robot pilot protocol covers the broader task-level evidence around this metrology layer.
Scope boundary: This guide is a requirements and acceptance framework, not a calibration certificate, ISO conformity statement or accuracy claim for any Matrix Dimension product. The actual device, task, site conditions, complete standards, measurement uncertainty and process tolerance must set project limits and revalidation intervals.
Frequently asked questions
Why can a highly repeatable robot still miss an offline-programmed point?
Repeatability describes how closely repeated outcomes agree, not how close they are to an external target. Offline paths also depend on robot kinematics and the base, TCP, work-object and CAD registrations; a stable error in any one can shift the path.
Is TCP calibration the same as absolute-accuracy calibration?
No. TCP calibration establishes the tool relative to the flange. Absolute-accuracy work can also identify robot kinematic parameters, while base, work-object and camera frames require their own registration and validation.
Can vision compensation replace robot calibration?
Sometimes it can correct observable part or target offsets, but camera calibration, occlusion, field of view and measurement uncertainty join the error chain. Offline paths, unseen motions and multi-robot coordination may still require arm and frame-registration evidence.
Sources
These primary sources support the material facts and engineering boundaries discussed above.
- ISO 9283:1998 — Manipulating industrial robots — Performance criteria and related test methods
- ISO/TR 13309:1995 — Test equipment and metrology methods for ISO 9283 evaluation
- ISO 9787:2013 — Robots and robotic devices — Coordinate systems and motion nomenclatures
- NIST — Efficiently Improving and Quantifying Robot Accuracy In Situ
- NISTIR 8093 — Foundations for Industrial Machine Registration
- NIST — Industrial Robot Accuracy Degradation Monitoring and Quick Health Assessment
Evaluating robot control, bimanual manipulation or a mobile platform?
Talk to Matrix Dimension →