Muskeology
Frontier tech, minus the hype

Robotics

Calibration Drift And Why Precision Fades

An arm that placed parts accurately on installation gradually stops doing so, because temperature, wear, collisions and tool changes all move the relationship between commanded and actual position.

Silhouette of wind turbines on a hill with a dramatic cloudy sky at sunset.
Silhouette of wind turbines on a hill with a dramatic cloudy sky at sunset. · Photo via Pexels

Repeatability is quoted in fractions of a millimeter, and cells are designed around it. What the specification does not say is that the number describes a machine in a known state.

Repeatability and accuracy are different claims

Repeatability is the ability to return to the same place it was taught. Accuracy is the ability to reach a place described in coordinates it was never physically shown.

Industrial arms are far better at the first than the second, which is why teaching by demonstration remains standard and why offline programs need touch-up on the floor.

Drift attacks both, but it is repeatability loss that operators notice first, because taught points stop landing.

Temperature moves the machine

Gearboxes and motors warm up over a shift, and metal expands. An arm taught cold will place slightly differently once it is hot, and the offset is not uniform across the envelope.

Plants sometimes see first-shift scrap that disappears by mid-morning and conclude the operator is at fault. The cause is a machine that has not reached thermal equilibrium.

Warm-up routines exist for this reason, and precision applications are often given a fixed period before quality parts are expected.

Wear and collisions change the geometry

Gear backlash grows slowly, belts stretch, and a minor collision can move a tool or bend a bracket by an amount too small to see and large enough to matter.

Because the controller reports joint angles rather than tool position, the arm continues to believe it is correct. Nothing in the fault log records the loss.

This is why any collision, however light, is a reason to re-verify a taught reference point rather than simply restart.

The tool is the part that changes most

Grippers are swapped, jaws wear, welding tips erode and suction cups compress. The arm has not moved, but the point that touches the work has.

Tool center point calibration exists to re-establish that relationship, and skipping it after a change is one of the most common causes of a cell that was fine yesterday.

Fixtures that let a technician verify the tool against a known reference make this a five-minute check rather than a debugging session.

Detecting drift before scrap does

The cheapest instrumentation is a reference feature the robot touches or images periodically and reports against. Deviation over weeks is visible long before parts fail inspection.

Plants that log this treat calibration as preventive maintenance with a schedule. Plants that do not treat it as an emergency, several times a year.

agricultureharvestingweedingautonomy
Tobias Nkemelu
AI & Compute, Muskeology

Tobias builds and breaks machine learning systems for a living, which makes him a difficult audience for benchmark announcements.

More from Tobias →

Also by Tobias Nkemelu

Robotics

Autonomous mobile robots outside the warehouse

Hospitals, hotels, factories and pavements — where wheeled autonomy has spread, and the specific reasons each environment is harder than a warehouse.

Lena Brandt··3 min read

Neurotech

Reading a neurotechnology paper

What to look for in a study, and which sections tell you whether the headline is supported.

Ravi Shankaran··3 min read