RCI 011 · Original data investigation

What 15 public product records publish—and omit.

The launch audit asks a narrow question: can public manufacturer evidence support useful comparison fields across actuation, perception, compute and power without flattening unlike conditions?

Published 2026-08-0911 min readDataset v0.1.0
25component records
25manufacturers
159claim records
79%selected comparison cells filled

A pressure test, not a market census

The sample contains three integrated or networked actuators, three robot stereo/depth cameras, three 3D LiDAR sensors, three embedded compute platforms and three battery or cell records. Products were selected to make unlike architectures collide with one schema—not to estimate market share or declare category leaders.

For each comparison, RCI fixed six decision fields before checking the records. A field counted as filled only when the selected official source supported a model-level value or bounded description. A filled cell does not mean the value is comparable; it only means a traceable public claim exists.

Interpretation limit

The percentages below describe this 15-product audit sample and selected source set as checked on 2026-08-09. They must not be generalized to every product from the same manufacturer or to the whole market.

Public records fill 75 of 90 selected cells

ComparisonRecordsFilled cellsCompletenessFields with at least one gap
Integrated joint actuator evidence table419 / 2479%Continuous / rated torque; Speed claim; Mass; Reduction ratio
Robot stereo and depth camera evidence table422 / 2492%Published range; Accuracy / precision
Robot 3D LiDAR evidence table314 / 1878%Channels; Power; Synchronization
Robot main-compute platform evidence table729 / 4269%CPU; AI acceleration claim; Memory; Camera I/O; Power envelope; Software baseline
Robot battery and cell evidence table419 / 2479%Energy; Discharge current; Pack communication; Safety / transport evidence
Robot USB-to-CAN interface evidence table212 / 1486%CAN FD support; Driver / API baseline

The actuator and depth-camera samples each fill 16 of 18 selected cells. LiDAR and compute each fill 14. The power sample fills 15. The apparent completeness is deceptive: many populated cells still use incompatible labels, operating modes or boundaries.

The larger problem is not missing numbers—it is missing equivalence

All three actuator records publish a headline torque and speed, but one headline is peak torque, one is stall torque and one is peak torque paired with a separate rated value. Only two selected records provide a continuous or rated field, and the visible source conditions are not identical. Converting every torque to N·m would produce a numerically tidy but technically false comparison.

The same pattern appears in perception. Camera records use ideal, optimal and product-mode language; LiDAR records distinguish reflectivity-conditioned range, instrument range and maximum range. Resolution and frame-rate maxima frequently belong to different modes. A product-family page may publish the highest channel count or point rate without proving that every orderable variant has it.

Compute fields are even more architecture-dependent. NVIDIA publishes an INT8 sparse TOPS figure. The selected Qualcomm brief describes heterogeneous accelerators but not a directly comparable TOPS figure. Raspberry Pi publishes CPU, memory, I/O and a long production commitment without a built-in NPU claim. “Unknown” in the TOPS cell is therefore a documentation result, not a zero.

All 89 launch claims still begin as manufacturer evidence

Evidence stateCountMeaning in this release
M · Manufacturer claimed157Transcribed from an official product page, manual or data sheet.
D · RCI derived1One nominal-energy calculation with formula and limitation shown.
V · Independently verified0No launch claim has yet passed the independent verification gate.
C · Conflicting1No formal conflict group is published yet.
U · Unknown0Unknowns are rendered as absent cells rather than synthetic claim rows in v0.1.0.

This is the most important result of the audit. Official manufacturer documents are appropriate sources for identity and reported specifications, but they do not independently validate performance. RCI therefore does not describe the launch catalog as verified hardware performance.

The next release should invest in conditions and integrations

Adding more model names would increase page count without solving the evidence problem. The next actuation sprint will prioritize duty-cycle curves, thermal assumptions, exact bus implementation, output sensing, bearing-load limits and revision identity. The non-joint tracks will prioritize camera/LiDAR test conditions, carrier-board sensor combinations, compute power modes, battery communication and pack-level safety evidence.

Independent evidence will be added claim by claim from reproducible papers, public BOMs, certification listings and disclosed laboratory tests. A third-party source will not be labeled verification merely because it repeats a vendor table.

  • Expand actuation from three to 25–50 audited records after the schema remains stable.
  • Add 5–10 further samples each for vision/LiDAR, compute and power before declaring their comparison fields mature.
  • Create explicit conflict groups when revisions or credible sources disagree.
  • Request factual review from manufacturers without granting editorial control.

Download and reproduce the audit

The complete product, manufacturer, claim, source, condition and comparison records are available in JSON and flattened CSV. The public methodology defines the evidence states and publication gate.

Suggested citation: Robot Component Index, “15 Robot Components Audited: What Public Product Records Publish — and Omit,” dataset v0.1.0, 2026-08-09.