128 files. 685,400 rows. One unresolved timing conflict.

The direct answer

Is the deposited dataset structurally intact? Yes within the checks RCI performed. The downloaded 25,857,481-byte ZIP matches the MD5 published by Zenodo. It contains 128 CSV files plus one README. RCI counted 685,400 data rows, found no blank, non-finite or malformed numeric rows, and reproduced all eight row totals printed in the paper’s Table 2.

Did RCI reproduce the paper’s neural-network accuracy? No. The article reports a strong same-motion result and a large direct cross-unit generalization gap. Those values are useful peer-reviewed source evidence, but the deposited package contains no training code, saved model, locked software environment or complete stochastic training specification. RCI labels every model metric below as paper-reported—not independently reproduced.

Is the sampling rate settled? No. The paper and README state 12 ms, approximately 83.3 Hz. The README also defines Sample as a controller timestamp in milliseconds, while every one of the 685,272 observed adjacent intervals advances by 4 units. Without the controller trace/export configuration or an author correction, neither 83.3 Hz nor 250 Hz should be presented as verified for the deposited rows.

Eight datasets reconcile exactly

R1 contributes 342,721 rows across D1–D4; R2 contributes 342,679 across D5–D8. Each robot has sixteen measured payload configurations and four recordings per payload. Corresponding masses differ by up to 3 g, so the labels are measured values—not a shared nominal class copied across both units.

The CSV contract is consistent across all files: six signed Iststrom motor-current columns in amperes plus Sample. RCI’s aggregate signed means differ between the two units on every axis; the largest absolute R2-minus-R1 mean difference is 0.129 A on A4. This is a bounded dataset observation, not a diagnosis of sensor drift, wear, efficiency or factory calibration.

One motion, two units, controlled mass geometry

The authors report two KUKA KR3 R540 six-axis robots, each driven by a KUKA KRC4 Compact controller running KSS 8.5.8. KUKA Trace recorded the NextGenDrive Actual Current channel for all six axes. A Python 3.11 application configured traces remotely and downloaded completed recordings over SFTP; this is cycle-level post-processing, not hard real-time feedback during motion.

Only A2 was commanded: from −75° to −120° and back, five times, at 15% of maximum robot speed and a reported peak near 80°/s. The other drives remained active. A pneumatic gripper held stackable discs from 434 g to 2609 g while geometry, gripping position, center of gravity and inertia distribution were kept approximately controlled.

That design isolates one useful question—whether a repeatable current signature contains mass information—but it does not test arbitrary pick-and-place trajectories, changing centers of mass, different acceleration profiles, another controller family or another robot model.

The public timing statements do not reconcile

If the README’s millisecond definition is applied literally, a four-unit step is 4 ms and the median recording spans 21.4 s. That conditional calculation is consistent with the roughly 20-second horizontal range shown for one representative trace in Figure 4, but visual consistency is not enough to overwrite the explicit 12 ms statement. The files could reflect an undocumented base tick, export rule or processing step.

What the paper reports about cross-unit transfer

The source paper extracts mean, minimum, maximum, median, standard deviation and signal energy from each axis, standardizes from the training data, and feeds the features to a 512–512 ReLU MLP trained for 150 epochs with Adam at a 0.001 learning rate. Table 5 reports:

The direct R1-to-R2 test is the important boundary: the reported MAE rises to 80.54 g and ±15 g accuracy falls to 14.60%, even though both units share a model and controller family. Mixed-unit training reports 5.37 g MAE and 96.88% within ±15 g on held-out D4 and D8. The paper attributes the gap to possible unit-specific mechanical, electrical or calibration differences; RCI does not turn that list of possible causes into measured diagnoses.

Those figures belong to the paper’s controlled motion. They are not product specifications for the KR3 R540, not a warranty of payload sensing, and not evidence that a production fleet needs one fixed recalibration interval.

What this changes for robot integration

1. “Same model” is not a calibration record. A data-driven virtual sensor must document which physical units supplied training data and how a new unit was normalized or calibrated.
2. Timing belongs in the evidence package. The trace source, controller time base, exported timestamp unit, effective row cadence and any downsampling must be independently checkable.
3. Split by recording, not adjacent rows. The paper’s leave-dataset-out design avoids leaking temporally correlated samples from one trace into both train and test sets.
4. Mass is not the whole payload state. Geometry, center of mass and inertia were controlled here. Varying them would be a different validation problem.
5. Post-cycle verification is not hard real time. The current implementation transfers files after a motion segment. Collision response or in-motion safety needs a different timing and safety case.

What RCI independently did

1. Resolved the concept DOI to Zenodo record 10.5281/zenodo.21456277 and verified the item-level CC BY 4.0 declaration.
2. Downloaded KUKA_payload_dataset.zip, matched Zenodo’s MD5 efe9394bf2579f380832eed6d501c614, and recorded SHA-256 51f93e8c453dd857170698627dc1d64ef14839e5c6efdb68148d01e0100c1253.
3. Parsed every CSV using the seven-column semicolon-delimited contract and rejected malformed, blank or non-finite numeric cells.
4. Derived robot, payload and dataset IDs from the documented filename contract; then counted files and rows independently rather than copying Table 2.
5. Compared D1–D8 totals with the paper and recorded eight exact matches.
6. Checked every first Sample value and all 685,272 adjacent differences, producing one observed step: four units.
7. Computed aggregate signed-current means only as a source-data characterization. No source row, paper figure or model output is redistributed.

The deterministic audit script is maintained in the RCI project as scripts/analyze-kuka-payload-dataset.mjs. It verifies the public release against the downloaded source package when both are available. The source archive itself remains on Zenodo.

Limits that stay attached

• RCI did not operate either physical robot, inspect the controllers or measure the payloads on an independent scale.
• The two unit labels and exact robot/controller/software identities are author-reported; the public archive contains CSVs and a README, not unit serial numbers or controller configuration exports.
• The timing conflict remains unresolved; duration-dependent metrics are conditional rather than verified.
• RCI did not reproduce the MLP results because the source package does not include training code, saved weights, a locked environment or all stochastic training settings.
• Aggregate signed currents are motion- and coordinate-specific and must not be used as cross-model performance rankings or motor-efficiency measurements.
• The paper's two-unit result does not estimate the distribution of unit-to-unit variability across a production fleet.
• RCI publishes only independently computed aggregates and source links; it does not mirror the third-party ZIP, raw CSV rows, paper figures or article text.

Source, attribution and rights

The paper is Bátrla et al., “Sensorless Payload Estimation in an Industrial Robot Using Internal Motor-Current Signals”, Sensors 2026, 26, 4867, published under CC BY 4.0. The source data are “Motor-Current Dataset for Cycle-Level Payload Estimation on Two KUKA KR3 Robots”, version DOI 10.5281/zenodo.21456277, also marked CC BY 4.0 by the depositors.

RCI’s JSON and CSV contain independently computed structural aggregates and source-reported model-result fields with explicit evidence roles. RCI does not mirror the third-party archive, raw rows, article text or figures, and does not relicense KUKA marks or other third-party material.

Download the audit

Download the immutable JSON release and dataset-level CSV. Stable aliases are current JSON and current CSV.

Suggested citation: Robot Component Index. KUKA Motor-Current Payload Dataset: 128 Files, 685,400 Rows, One Timing Conflict. RCI 026, research dataset v0.1.0, 2026-08-10. https://robotcomponentindex.com/research/kuka-motor-current-payload-dataset-audit/