🧬 DriveDNA-Controlled — 14 drivers, one Honda Civic, the same Tampa roads

_{All 14 drivers, one frame each — the same car, different people behind the wheel. 11 of them are shown at the moment they passed one shared location on the core road network; `driver_467` and `driver_468` at other points of the same network; `driver_472` elsewhere in the area. Frames are taken from the released forward-camera videos; day and night sessions are included.}

DriveDNA-Controlled is a local naturalistic driving dataset recorded in Tampa, Florida, on the public roads around the University of South Florida (USF) campus, released as an additional, controlled supplement to DriveDNA. Fourteen participants drove the same Honda Civic (a single vehicle, platform HONDA_CIVIC) during their own everyday, commute-style driving in the area, and most drives pass through the same core road segments; all fourteen consented to public release and are included here, and none of them appears in the main release. Vehicle and road are therefore held approximately constant while the driver changes, which is exactly the comparison the main corpus — one car per driver, drivers spread across continents — cannot offer.

56 drives · 14 drivers · 9.7 h recorded · 312 min of moving human driving (is_human, vEgo > 2 m/s) · 611 benchmark windows · 2.0 GB

Layout, column schemas, the pseudonym scheme and the 10 Hz signal format follow the main release, so loaders written for DriveDNA work on this folder. Driver ids continue the main sequence: the 14 drivers are driver_461 through driver_474, so the two releases can be concatenated without collisions.

---

Experimental design

Location	Tampa, Florida, USA — public roads around the University of South Florida campus. As in the main release, the tables contain no GPS coordinates; road geometry is present only through the driving signals themselves (actual_curvature, curv_measured, lane offsets) — the same 27 columns as DriveDNA
Vehicle	One physical Honda Civic 2020, driven by every participant. model_canon is HONDA_CIVIC in every table — the value the main release uses for this car
Drivers	14 volunteer participants, all of whom consented to public release and are included here as driver_461 through driver_474 (14 consecutive ids continuing the main release's sequence); none of them appears in the main DriveDNA release
Sessions	September 2026. The 14 drivers carried out the experiment and had their driving trajectories recorded. Each recording is released as a separate drive of its driver, numbered chronologically
Roads	Everyday driving on the street network around USF. Most drives traverse the same core road segments (the figure above shows one of them), so road geometry, speed limits and traffic environment are largely shared across drivers; a minority of drives also cover other nearby streets. It is naturalistic driving, not a fixed lap on a closed course

Per-driver summary

Driver	Drives	Minutes recorded	Human-moving min	Windows
driver_461	23	249.5	133.2	241
driver_462	2	42.5	23.8	54
driver_463	2	25.4	10.6	22
driver_464	2	17.7	11.1	22
driver_465	3	31.1	19.2	41
driver_466	2	16.9	10.4	22
driver_467	2	27.1	18.3	38
driver_468	2	12.4	8.1	15
driver_469	2	14.0	9.0	18
driver_470	3	26.9	15.0	31
driver_471	7	37.3	23.7	44
driver_472	4	52.1	12.2	24
driver_473	1	12.9	5.5	11
driver_474	1	17.1	12.1	28

Window scenarios: stop_go 259 · curve 258 · car_following 40 · urban 39 · free 15.

---

📁 Layout (mirrors the main release)

index/drives.parquet|csv           one row per drive — main-release columns + supplement columns (below)
index/human_segments.parquet|csv   87 human-driving spans (i0/i1 row range, t0/t1 s, k_min = first segment number)
raw_signals/driver_XXX/drive_YYY.parquet   27-column 10 Hz table, same columns/dtypes as DriveDNA raw_signals/
raw_signals_csv/driver_XXX/drive_YYY.csv   the same tables as CSV
videos/driver_XXX/drive_YYY/segN.mp4       forward qcamera, 526×330, nominal 20 fps, stream-copied with original timestamps, audio removed (606 files)
data/windows.parquet               611 windows: 12 descriptors + 6-way scenario + 8 behavioural primitives
features/windows_x.npy             (611, 600, 17) float16 — 17 channels × 60 s @ 10 Hz (channels.json), windows_mask.npy
features/windows_vid*.npy          window-aligned frozen video features (DINOv2 / DINOv3 / SigLIP2 / V-JEPA 2)
embeddings/<family>/driver_XXX/drive_YYY.npz   per-segment video embeddings keyed segN
splits/supplement_drivers.json     the 14 driver ids and their drives

Supplement-specific index columns

index/drives starts with the ten main-release columns, then adds:

Column	Meaning
seg_first, seg_last, n_segments	original one-minute segment numbers covered by this drive
n_video_segments	qcamera segments released for the drive
duration_s	table duration in seconds

fold is supplement for every drive: these drivers were never in any DriveDNA training, validation, test or few-shot fold, so frozen models from the main release can be evaluated here as true unseen drivers.

Behavioural primitives

The eight p_* labels use the main corpus' per-scenario Q20/Q80 thresholds (not thresholds recomputed on this small set), so a label here means the same as in DriveDNA. The thresholds were verified to reproduce the stored labels of all 62,674 corpus windows before being applied.

---

Relation to DriveDNA

DriveDNA's benchmark asks whether a representation recognises the driver rather than the car, the route or the traffic situation. In the main corpus those factors are naturally entangled. This supplement supplies the complementary, controlled sample: the same car and largely the same roads for everyone, so a model that separates these drivers is separating people, not platforms or places. All 14 drivers were held out of every DriveDNA training, validation, test and few-shot fold, so frozen models from the main release can be evaluated here as true unseen drivers.

Known data notes

• The last one-minute segment of most drives is partial. 3 in-range segments (3 drives) have no video because their qcamera file was empty (n_video_segments < n_segments); four trailing log stubs without vehicle data are excluded from n_segments.
• In 26 drives the first video segment holds fewer than 1,200 frames: 19 of them have one internal camera gap of 2–18 s starting about 10 s after the first frame (encoder start-up; 6 gaps exceed 10 s), 6 end a few frames early with no internal gap, and the rest are short single-segment drives. Timestamps are preserved, so frames stay aligned to the table; the 2 fps video features repeat the last available frame across a gap.
• Table time 0 is the first vehicle-bus sample of the first segment, which can lag the first video frame by a few seconds (the same convention as the main release).
• gas and yawRate are identically 0 in every table: these openpilot 0.11.1 Honda logs do not report pedal position or a CAN yaw rate (gasPressed, brake, brakePressed and the localizer yaw_rate are populated). This follows the main-release convention for unavailable Tier-B signals.
• Segment numbers follow the original recordings, so a drive can start above segment 0 (seg_first, k_min) when earlier segments of that recording are not part of the drive.
• The 10 Hz grids are gap-free: no drive has a carState gap longer than 0.2 s, so no stretch of any table is interpolated across missing data.
• Clocks were aligned to GNSS time before deriving year_month and the chronological drive numbering (one drive without a GNSS fix used the synchronised system clock).

🔐 Privacy

• Drivers appear only as driver_XXX; the VIN and the private name mapping are held by the authors.
• No GPS, no timestamps finer than the month (year_month, as in the main release), no cabin or wide camera, no audio (microphone tracks present in the source files were removed).
• Only the low-resolution forward qcamera view is released (526×330); the high-resolution road camera, which can resolve licence plates and faces of other road users, is withheld.

Citation

Please cite the DriveDNA paper (arXiv:2607.23822) and refer to this release as DriveDNA-Controlled.