Two 4.3-hour traces. 310,697 BMS rows. One unloaded boundary.

The direct answer

Are the public clean windows complete on their stated grid? Yes. RCI independently found 155,924 condition-A rows and 154,773 condition-B rows, no missing BMS state inside either window and exactly 0.1 seconds between every adjacent row.

Does “10 Hz” mean ten new high-resolution BMS measurements per second? No such conclusion is available. The package documents a 10 Hz controller polling / recording cadence and a 10 Hz public zero-order-hold grid, but no BMS response counter, response timestamp or internal sensing bandwidth. Stored pack current lies on a 0.1 A grid and voltage on a 0.5 V grid. In both clean windows, roughly 97.3% of adjacent rows carry the same two stored pack-current values.

Does more motion coincide with more reported current? In condition B, yes descriptively: the highest activity quartile averaged 8.17 A versus 6.39 A in the lowest quartile, a difference of 1.786 A or 28.0%. The relationship remains positive after removing each minute’s mean. Condition A does not show the same within-minute pattern.

Is that actuator energy or a battery benchmark? No. It is a same-clock association during two unloaded scripted routines. It does not separate motors from compute, sensors, fans, controllers and other loads; it does not compensate for BMS response timing; and it contains no independent current or voltage reference.

Two records, kept separate

RCI does not pool the records. The source identifies the same operator-provided model in both, but does not establish that the physical robot or either pack is the same unit across dates. Run-local pack indices and SOH fields therefore stay run-local.

A 10 Hz table can still contain long runs of unchanged values

Both BMS tables are perfectly regular at the public 100 ms grid. That is a useful integrity result—but not a measurement-resolution result. RCI found current-pair transitions only 4,238 times in condition A and 4,038 times in condition B. The median exact-equality run was 3.5 seconds in both conditions; the p95 run was 7.0 seconds.

The public state_age_s field tells a related story: its median is 1.8 seconds in both windows, and 70.4% of condition-A rows and 69.4% of condition-B rows carry a retained state first seen more than one second earlier on the public grid.

Condition B has the stronger motion association

RCI marks a joint channel active when its p99-minus-p01 clean-window position range is at least 0.05 rad. For each 100 ms step, the activity index sums absolute position changes across those active radian channels and divides by 0.1 seconds. Only complete ten-row one-second bins enter the association analysis.

Condition A: 15,591 complete seconds, fourteen active arm channels, same-second Pearson 0.095 and Spearman 0.039. The global high-minus-low difference is only 0.208 A; after de-meaning within each minute, Pearson is effectively zero (0.00002), and the median minute-level high-minus-low difference is -0.035 A.

Condition B: 15,474 complete seconds, nineteen active joint channels, same-second Pearson 0.417 and Spearman 0.463. Within-minute de-meaned Pearson remains 0.377. Across 257 eligible minutes, the median high-minus-low current difference is 1.365 A; its minute-level interquartile range is 0.666 to 2.250 A.

The threshold sensitivity is unusually clean. At active-channel thresholds of 0.01, 0.05 and 0.10 rad, condition B’s global high-minus-low difference remains between 1.786 and 1.786 A. The conclusion is not an artifact of the selected 0.05 rad cutoff.

RCI does not search time lags for a larger coefficient. The dataset lacks a distinct BMS response timestamp, and selecting the best lag after seeing the result would add an avoidable researcher degree of freedom.

Battery fields: useful aggregates, not calibrated energy

Direct 0.1-second integration of the reported total current gives 30.166 Ah for condition A and 30.737 Ah for condition B. Multiplying each pack’s BMS-reported current and voltage before integration gives 1,462.0 Wh and 1,487.8 Wh.

Those Wh values are field-product estimates—not independently metered energy. The source metadata describe two 48 V, 17 Ah LFP packs in simple parallel for each run, but pack and BMS manufacturer/model, age, cycle history, ambient conditions and external reference instruments are unavailable.

The SOC endpoints also remain reported metadata: condition A goes from 100% / 100% to 0% / 7.6%; condition B goes from 100% / 97.6% to 2% / 0%. They are not true-SOC calibration points. No cell voltages or independent capacity measurements are present.

What RCI independently recomputed

1. Archive identity: downloaded only robobatt_10hz_release_v1.0.0.zip, verified 185,123,569 bytes and SHA-256 f3b98405f9689f35a740da38a7f6626416328d68dc1c040bf01fee14cf94de1a.
2. Inner integrity: recomputed all 138 listed file sizes and hashes. None were missing or mismatched. The three unlisted files are the manifest itself and two validation reports.
3. Upstream contract: reran the deposited validator and obtained all_checks_pass: true. That confirms the source package reproduces its own contract; it is not the RCI result.
4. Independent aggregates: read the public BMS and latest-message joint/chassis tables, reusing only the source-declared clean-window timestamps. RCI did not use upstream derived summary tables as numeric inputs.
5. Current and energy: integrated at 0.1 seconds and checked that stored total current equals the two pack currents within the public rows.
6. Motion association: used complete one-second bins, global activity quartiles, Pearson and rank correlation, within-minute de-meaning, within-minute high/low comparisons and three active-channel thresholds.

What this release does not establish

• No payload performance, gait efficiency, maximum runtime, range or production duty-cycle result.
• No causal allocation of battery current to a joint, actuator, chassis subsystem or controller.
• No internal BMS sampling bandwidth, response latency, accuracy or protocol result.
• No cell-level voltage, temperature, balance or fault evidence.
• No independent current, voltage, SOC, SOH, temperature or energy calibration.
• No cross-condition ageing conclusion; physical robot and pack continuity are unknown.
• No reuse of raw source rows, native HDF5 streams, publication plots, PDFs or photographs by RCI.

Source, attribution and rights

Source dataset: Guo, Feng and Liu, Hongxing. RoboBATT: Multirate Robot Battery and Actuator Telemetry Traces, version v1.0.0. Zenodo, published 2026-08-08. Version DOI: 10.5281/zenodo.21853137. The source data and non-code materials are declared CC BY 4.0; deposited code is MIT.

Changes made by RCI: verified the lightweight archive and its internal manifest; selected and disclosed an aggregate activity definition; recomputed condition-level statistics, correlations and sensitivity results; wrote new interpretation and limitation text; and published only aggregate JSON/CSV. RCI does not claim authorship of RoboBATT or relicense unrelated third-party names and marks.

Download and cite

Download the immutable aggregate JSON and condition-level CSV. Stable aliases are current JSON and current CSV.

Suggested RCI citation: Robot Component Index. Robot Battery and Actuator Telemetry: Two 4.3-Hour Traces, One Unloaded Boundary. RCI 025, research dataset v0.1.0, 2026-08-10. https://robotcomponentindex.com/research/robot-battery-actuator-telemetry-audit/

Source citation: Guo, Feng and Liu, Hongxing. “RoboBATT: Multirate Robot Battery and Actuator Telemetry Traces.” Version v1.0.0, Zenodo, 2026-08-08. DOI: https://doi.org/10.5281/zenodo.21853137.

Use the RCI methodology for evidence-state definitions, the battery compatibility guide for system boundaries and Cite RCI for immutable citation records.