Deployability reporting audit dataset for "Edge-Device Deployable Self-Supervised Traversability Estimation for Field Robots: A Critical Review"

Companion dataset for the review article "Edge-Device Deployable Self-SupervisedTraversability Estimation for Field Robots: A Critical Review". WHAT THIS IS Per-paper coded indicators from a structured audit of how often the field-robotics perception literature reports the measurements that decide whether a method can run on edge hardware. The audit covers 225 unique studies, appearing 249 times across three thematic populations: segmentation and quantization for edge NPUs (79), self-supervised label generation for traversability (73), and multi-modal sensor fusion under embedded compute (97). Each paper was scored on four binary indicators: whether it reports a numeric power envelope in watts, whether it reports its own observation of thermal behaviour under sustained load, whether it reports a numeric quantization-induced accuracy change measured on the same device, and whether it reports inference latency on named edge hardware. HEADLINE FINDING Across the 225 unique papers: Power envelope 17 / 225 = 7.6% Thermal behaviour under sustained load 4 / 225 = 1.8% Quantization-induced accuracy change 29 / 225 = 12.9% Edge-hardware inference latency 62 / 225 = 27.6% All four reported 0 / 225 = 0.0% No paper in the audited corpus reports all four deployment quantities. CODING VALIDATION A random 12% sample (30 of 249 appearances, stratified by theme, random seed 20260815) was independently re-coded blind against the same criteria. Raw agreement across the four indicators was 94.0% (Cohen's kappa 0.69; prevalence-adjusted kappa 0.88). All disagreements were adjudicated against the source papers. The sample, the independentre-codings, the original codes, the per-indicator agreement statistics, and the full list of adjudicated corrections are included in the five files prefixed validation. WHAT CHANGED IN VERSION 3 Version 3 is the first release in which coded indicator values change. Versions 1 and 2 carried identical codings; version 2 added bibliographic columns only. Following the blind re-coding validation, nine indicator values across seven papers were corrected. Four corrections lowered a reported rate and three raised one, so the errors were not systematically directional. Each headline rate moved by less than one percentage point and the relative order of the four categories did not change. The most consequential change is that both papers previously coded as reporting all four indicators were re-read in full and found to be over-coded. One substituted a configured platform power mode for a measured power draw; the other gave a speculative thermal remark and an accuracy table not separated by numeric precision. Users reproducing statistics from versions 1 or 2 should note that those releases carry the pre-correction codings and will not match the published article. FILES meta_theme2.csv, meta_theme3.csv, meta_theme4.csv Per-paper codings, one file per theme, with bibliographic columns. meta_extraction_summary.md Aggregate counts, per-theme breakdowns, confidence intervals, significance tests. README.txt Full column definitions, coding rules, and limitations. validation_*.csv (five files) The blind re-coding sample, the original withheld codes, the independent re-codings, per-indicator agreement statistics, and the adjudicated corrections. LICENSE Creative Commons Attribution 4.0 International (CC-BY-4.0). CITATION Please cite both the article and this dataset. Use the concept DOI 10.5281/zenodo.20586733, which always resolves to the most recent version.