A warehouse robot begins missing pallet pockets by twelve millimeters. The navigation stack is unchanged. The camera still streams. Every health check is green.
Three shifts later, someone notices that the camera bracket was replaced during routine service. The replacement moved the optical frame slightly. A technician ran calibration, copied a YAML file onto the robot, and closed the ticket. Nobody can prove which file is active.
The robot does not have a calibration problem. The product has a calibration custody problem.
A calibration file is not a product state
Calibration turns physical observations into parameters that software can use. OpenCV, for example, models camera intrinsics, radial distortion, tangential distortion, and the relationship between pixels and real-world units. Its workflow can also save results and calculate reprojection error.
Those numbers look authoritative when they sit in a configuration file. Yet the file rarely answers the questions that matter during a field failure. The team still needs the sensor identity, lens, resolution, target, software version, temperature, mounting state, and approving person.
Treating calibration as a loose file creates what I call parameter amnesia. The system remembers the coefficients and forgets the conditions that made them valid.
ROS calibration parsers show the useful mechanical layer. They can read named calibration data, image dimensions, projection values, and distortion models into a camera information message. That gets parameters into the runtime. It does not establish why this particular robot should trust this particular record today.
Give each result a chain of custody
A calibration ledger should make every active parameter set a versioned product object. The object needs an identity, a history, and a defined scope. It should be possible to move from a strange robot behavior to the exact evidence behind the active calibration without asking who last touched the machine.
At minimum, each ledger record should contain:
- Asset, sensor, lens, mount, and robot identifiers
- Calibration method, target, tool version, and operator
- Input dataset or durable evidence reference
- Output parameters, units, uncertainty, and quality scores
- Environmental conditions and applicable operating range
- Effective time, superseded record, reviewer, and reason for change
NIST defines metrological traceability as a documented, unbroken chain of calibrations, with each step contributing to measurement uncertainty. A warehouse camera calibration does not automatically require a national metrology program. The product lesson still travels well. A number is more trustworthy when its origin, uncertainty, and fitness for the intended use remain visible.
That last distinction matters. NIST also notes that traceability alone does not guarantee fitness for purpose. A technically valid calibration can still be wrong for a robot that changed resolution, focus, payload, mounting geometry, or operating temperature.
Make calibration validity executable
The ledger becomes useful when the runtime can evaluate it. Each robot should advertise the calibration record identifiers currently in force. Telemetry and incident bundles should include them. A deployment should refuse or degrade operation when required records are missing, expired, or incompatible with the installed hardware.
Validity rules should follow the product boundary. A camera record may bind to sensor serial number, lens, resolution, binning mode, focus setting, and mount revision. A force sensor may bind to tool mass, orientation, and temperature range. An arm may need separate kinematic records after joint or end-effector service.
Teams already make a similar architectural choice when they decide how a camera interface sets the product boundary. Calibration deserves the same attention because it connects hardware identity to software meaning.
Do not reduce validity to a calendar reminder. Time can be one trigger, but physical events are often better. Recalibration may be required after a collision, bracket replacement, firmware change, lens swap, thermal excursion, or quality drift. The ledger should encode those triggers and show which assets became suspect.
Design the field workflow before the database
The hard part is not storing another document. It is making the correct behavior easier than the shortcut.
A technician needs a guided procedure that identifies the asset, checks prerequisites, captures evidence, calculates quality metrics, and proposes a new record. A reviewer needs the old and new values, the expected impact, and a clear reason for acceptance. The robot needs a signed or otherwise protected reference that it can verify at startup.
Rollout should be reversible. A new record can be staged on one robot, compared against known tasks, and then promoted across a fleet. If performance worsens, the team should restore the prior record without erasing the failed attempt. Failed calibration is evidence too.
Product metrics should expose the custody gap. Track robots with unknown active records, calibrations outside their declared envelope, service events without required recalibration, and incidents that lack calibration identifiers. Measure the time from suspected drift to verified correction.
The useful promise is simple: every physical result can name the calibration state that shaped it. When that promise holds, a mysterious twelve-millimeter miss becomes an inspectable change instead of a three-shift investigation.
Comments
Comments are powered by GitHub Discussions through Giscus.