Calibration results and limits

The August 12 bundle remains the demonstrated manipulation reference. Later collections helped investigate the model’s errors but did not supply a deployed replacement.

Use pixel residuals to compare declared models on the same retained data and split. Use capture/cleanup status to judge the collection, and separately observed task outcomes to judge manipulation. These three results answer different questions.

How to read the reported errors

A reprojection residual is the distance in pixels between an observed marker corner and the corner predicted by the declared camera/arm model. Radial pixel RMS combines both image axes: sqrt(mean(du² + dv²)); it is not a direct millimetre error at the hand.

Training error measures the samples used to fit parameters. Grouped cross-validation (CV) fits without a held pose group and evaluates that group; frames from the same stationary burst stay together. Active-pose error isolates the excited arm configurations instead of averaging them with easier stationary anchors. Compare the same metric, groups and model freedom. A lower number from a different session or a more flexible model is not sufficient evidence of a better deployed calibration.

The August 12 stacking baseline

The original bilateral dataset contained 41 left and 62 right observations. The selected model fitted a shared camera, two target transforms and seven joint offsets, for 25 parameters. Its grouped cross-validation error was 5.409 px overall: 5.162 left and 5.566 right, with 4.952 px training error.

The bundle is included in demo main. It was used by the five retained August 25 demo runs and the earlier 13 completed August 21 stack requests. That is a useful deployment connection, but 13 completed recordings are not 13 independently scored stacking successes. The five August 25 completion records also need separate physical outcome annotation. The bundle does not establish absolute FK accuracy everywhere in the workspace.

Early effective corrections

The manual session contained 39 observations and 273 images. A camera/target extrinsic fit gave approximately 8.205 px residual. Adding an effective shoulder-roll correction of +4.346° reduced it to 4.875 px; allowing all offsets gave 4.363 px. The extra parameters did not by themselves identify a physical cause. Capture 028 was later confirmed to involve stream lag, not demonstrated compliance. Removing it barely changed the shoulder correction.

In a right-hand 62-primary study, fixing the marker at its CAD transform gave 23.045 px holdout error versus 9.737 px when it was free. The effective marker correction was 13.638 mm and 4.621°. Those are fitted compensation values, not measurements of how far the printed mount was misplaced.

September 5: broader capture, export still blocked

Session 20260905T231145 accepted 55 captures: 23 left, 24 right and eight anchors, totaling 385 images and 48 pose groups. It recorded 17 rejected attempts and 19 retries, then returned through Q and released without a fault. Its six-topic standing bag omitted the camera streams; images of failed views cannot be recovered from that recording.

Model on this session	Grouped CV, overall	Important qualification
Camera and targets	22.95 px	Native FK did not explain the full route
Shoulder correction model	5.94 px	Effective correction
Selected seven offsets	5.43 px	Active-pose error remained 7.95 px

Stationary observations were around 2 px. Mixing them with excited arm poses improves the overall number. The numerical similarity between 5.43 px here and August’s 5.409 px is not an equal-condition comparison.

The normal exporter rejected apparent anchor PnP translation spans of about 19.8 mm left and 10.65 mm right. Pooling all frames and both targets reduced the joint estimate to about 3.94 mm, and synthetic noise checks showed the original gate was uncertain. Neither analysis proves 20 mm physical camera drift or establishes a validated replacement acceptance gate.

September 7: complete images, unsuccessful cleanup

Session 20260907T133509 accepted 49 holds, yielding 343 images and 686 marker observations. Its 14-topic bag contained 2,298,946 messages and approximately 28.43 GB. All captures were accepted, but finger restoration failed: left index-proximal residual was 0.092489 rad against a 0.08 rad criterion after 576 commands over six seconds. Cleanup entered Damp. This was not a normal completed release, and it motivated the manually preclosed workflow.

Repeated configurations were consistent after measured-state compensation, with residuals of 0.107–0.361 px. Opposite-approach raw differences of roughly 3.71–16 px reduced substantially after accounting for measured configuration. However, all 12 A/B pairs failed the measured-FK pose-equivalence gates. The experiment therefore cannot isolate a cause at an identical hand pose.

The camera was deliberately moved between sessions. Measured joints, separate camera registrations and the operator’s corrections are required when interpreting the comparison. Command tracking error is normal control behavior; calibration already uses measured state.

What remains established

The August effective bundle supported the documented manipulation baseline. The September work showed that nominal geometry, parameter compensation and validation design matter. It did not produce a newly deployed bundle or a unique physical diagnosis. The geometry investigation explains why a plausible 5 mm wrist correction needs careful interpretation.

Evidence: dated tabletop log entries, original bundle evaluation and the calibration investigation ledger. Source identities.