HenryYHW/ADAS-TO_longitudinal

ADAS-TO longitudinal — closing case-control windows (n = 5090)

Companion subset of the ADAS-TO dataset for longitudinal (car-following) take-over modeling. This is a case-control set, not 5090 takeovers: 2293 takeover cases (arm T) and 2797 non-takeover closing controls (arm Q) drawn from the same drives.

Examples

Forward video (red flash = ADAS disengagement, takeover at 10 s) with the synchronized radar/CAN signals of the same clip.

Fast closing on a slower lead (~9 m/s approach) — TOYOTA_PRIUS/driver_142/route_001/0

Typical low-speed car-following takeover — TOYOTA_PRIUS/driver_182/route_010/2

Closing resolved to standstill — VOLKSWAGEN_SHARAN_MK2/driver_289/route_007/1

Case vs same-drive matched control — both contain sustained closing; only the case ends in a takeover:

Windows (windows.parquet)

Per window: 41 samples at 10 Hz, t in [-3.0, +1.0] s; t = 0 is the ADAS disengagement (cases) or the pseudo-event anchor (controls).

column	description
win_id	window id (T_* case / Q_* control)
t	time relative to (pseudo-)event, s
s	radar lead gap, m (NaN when no valid lead)
vr	radar relative speed vRel, m/s (closing speed = -vr)
vE	ego speed, m/s
aE	ego measured acceleration, m/s^2 (realized vehicle response)
valid	lead-valid flag (isfinite(s) & s > 0.5)

Meta (meta.parquet)

arm (T/Q), anonymized clip (cases; matches the public ADAS-TO release paths), car_model, anonymized driver/route, window mean ego speed v_mean, sustained-closing duration sus (s), maxDv/minTTC over the pre-event closing window, control sub-window offset t0_sub, and the case-control pairing (has_ctrl, matched_case, match_relaxed).

Clips (clips/)

For each case (arm T) the original 20 s clip is included at clips/{meta.clip}/: the forward video takeover.mp4 (takeover at 10 s) and 13 openpilot log CSVs (carState, carControl, carOutput, controlsState, drivingModelData, longitudinalPlan, radarState, accelerometer, VehicleIMU, Gyroscope, CameraOdometry, LiveCalibration, LiveParameters) plus a sanitized meta.json (anonymized driver/route ids). File layout and anonymization follow the main ADAS-TO release. Controls (arm Q) have no clips by construction — they are unlabeled engaged stretches of the same drives; their radar/ego series are in windows.parquet.

Construction

• Cases: all-corpus longitudinal takeovers with radar lead coverage >= 0.9 over video [7, 11] s and a sustained closing episode (approach speed

  = 0.5 m/s for >= 1.0 s) inside [7, 10] s (takeover at 10 s).

• Controls: same-drive engaged windows with no driver input and no disengagement, >= 60 s away from every takeover on the route, containing a closing episode under the identical criterion; pseudo-event = episode sub-window start + 3 s; speed-matched (+-5 m/s) per case, K = 2.
• 1,436 cases have same-drive matched controls; 857 have none (kept).

Caveats

• Case-control sampling: model outputs on this set are not absolute naturalistic takeover probabilities.
• aE after driver input may reflect the driver, not the automation.
• Radar-missed leads and cut-ins are out of scope by construction.