🚗💨 ADAS-TO-Sample

Open subset of ADAS-TO: six drivers, every clip

1,342 real-world takeover events · 16 vehicle models · 7 manufacturers · 15 files per clip

---

📢 What this repository is. ADAS-TO-Sample contains every clip of six anonymized drivers from the full ADAS-TO corpus. Paths, file schema and anonymization are identical to the full dataset: any clip here exists at the same path in HenryYHW/ADAS-TO. For the full dataset (16,446 clips, ~41 GB) request access at HenryYHW/ADAS-TO or contact henryyuhangwang@gmail.com. Paper: arXiv:2603.06986.

---

🎬 Takeover Examples

Each GIF shows ±3 seconds around the takeover moment — ADAS engaged → driver takes control

On-coming Traffic	Bridge	Night Driving	Sharp Curve
Surrounding Car	Traffic Light	Lane Change	Hard Brake

---

📊 Dataset at a Glance

	Statistic	Value
🎥	Takeover clips	1,342
🧑‍✈️	Drivers	6 (anonymized driver_124, driver_232, driver_251, driver_277, driver_347, driver_348)
🛣️	Routes	311
🚘	Vehicle models	16
🏭	Manufacturers	7 (Ford, Honda, Hyundai, Kia, Tesla, Toyota, Volkswagen)
⏱️	Clip duration	20 seconds (±10 s around the takeover)
📹	Video	Front-facing camera, 20 fps
📡	CAN / sensor logs	100 Hz (rlog, 679 clips) or 10 Hz (qlog, 663 clips)
📁	Files per clip	15 (1 video + 1 meta + 13 CSV)
💾	Size	~3.7 GB (20,130 files)
📦	Full dataset	16,446 clips · 364 drivers · 179 vehicle models · ~41 GB (ADAS-TO)

Clips per vehicle model

Vehicle model	Clips	Vehicle model	Clips
HONDA_CIVIC	366	HONDA_ACCORD_HYBRID_2018	65
TESLA_MODEL_X	118	TOYOTA_CAMRY_TSS2	64
TESLA_AP3_MODEL_3	117	VOLKSWAGEN_TIGUAN_MK2	53
HYUNDAI_IONIQ_5	112	KIA_EV6	42
TOYOTA_RAV4_TSS2_2023	108	KiaNiro2023	33
TOYOTA_CAMRY_2021	88	TOYOTA_RAV4_2023	23
FORD_MUSTANG_MACH_E_MK1	72	KIA_NIRO_EV_2ND_GEN	9
HYUNDAI_IONIQ_5_2022	71	FORD_MAVERICK_MK1	1

---

🔍 How the subset was defined

• The subset is defined by driver, not by random sampling: it holds all corpus clips of the six drivers listed above and nothing else.
• Because most of these drivers used one device across several vehicles, the subset spans 16 vehicle models, but its vehicle mix is narrower than the full corpus (179 models). Use the full dataset for cross-platform statistics.
• Anonymization, directory layout and file schema are the same as in the full dataset. The mapping from driver_NNN / route_MMM to real device or route IDs is not released.

---

📁 Dataset Structure

ADAS-TO-Sample/
├── <CAR_MODEL>/                        # e.g. HONDA_CIVIC, TESLA_AP3_MODEL_3
│   └── <driver_NNN>/                   # 🔒 anonymized driver ID
│       └── <route_MMM>/                # 🔒 anonymized route ID
│           └── <clip_id>/              # integer, 0-indexed within the route
│               ├── 🎥 takeover.mp4          20 s front-view video (20 fps)
│               ├── 📋 meta.json             clip metadata & timing
│               ├── 🚗 carState.csv          speed, accel, steering, pedals, cruise state
│               ├── 🎮 carControl.csv        lateral/longitudinal commands
│               ├── ⚙️ carOutput.csv          actuator outputs
│               ├── 🤖 controlsState.csv     ADAS controller state & alerts
│               ├── 🧠 drivingModelData.csv  lane-line estimates, desired curvature
│               ├── 📏 longitudinalPlan.csv  planner targets, FCW
│               ├── 📡 radarState.csv        lead-vehicle radar tracks
│               ├── 📐 accelerometer.csv     IMU acceleration
│               ├── 🧭 VehicleIMU.csv        body-frame IMU / yaw rate
│               ├── 🌀 Gyroscope.csv         angular rates
│               ├── 📷 CameraOdometry.csv    visual odometry
│               ├── 🎯 LiveCalibration.csv   device → vehicle frame calibration
│               └── 🔧 LiveParameters.csv    online vehicle-parameter estimates
└── annotations/
    └── clip_final_labels.csv           # scenario label per clip (see below)

15 files per clip (1 video + 1 metadata + 13 CSVs). Vehicle logs are sampled at 100 Hz where an rlog was available (meta.json: log_kind), otherwise at the 10 Hz qlog rate; video is 20 fps.

---

🏷️ Annotations

annotations/clip_final_labels.csv has one row per clip (1,342 rows) with the anonymized clip_path, car_model, clip_id, the scenario label final_label, lc_direction (left/right for lane changes), human_labeled / human_label (manual review status) and realign_status.

final_label	n	Meaning
cover	651	Maneuver-Filtered ADAS — ordinary lane-keeping / car-following, no maneuver confound
lane_change	252	lane transition, merge or fork (lc_direction gives left/right)
turn	220	intersection turn or departure from the through path
stop	194	stopping / decelerating for a traffic control
Unknown/other	25	complex geometry, non-routine scene, or insufficient evidence

Most analyses should start from cover, which removes disengagements that are explained by a planned maneuver rather than by the automation reaching its limits.

---

📐 Takeover Event Definition

  ◄──────── 10 seconds ────────►◄──────── 10 seconds ────────►
  ┌──────────────────────────────┬──────────────────────────────┐
  │      🤖 ADAS ENGAGED         │      👤 MANUAL CONTROL        │
  │   (automation driving)       │   (driver takes over)        │
  └──────────────────────────────┴──────────────────────────────┘
                                 ▲
                            TAKEOVER EVENT
                         (ON → OFF transition)

A takeover event is detected as an ADAS ON → OFF transition satisfying:

Criterion	Value
ADAS engaged	controlsState.enabled OR cruiseState.enabled
Min ON duration	≥ 2 seconds before disengagement
Min OFF duration	≥ 2 seconds after disengagement
Gap merging	Transient gaps < 0.5 s merged (filters sensor noise)
Clip window	±10 seconds centered on the transition (20 s total)

---

📑 Data Fields Reference

📋 meta.json — Clip Metadata

Field	Type	Description
car_model	string	Vehicle model (e.g., HONDA_CIVIC)
dongle_id	string	Anonymized driver ID (driver_NNN)
route_id	string	Anonymized route ID (route_MMM)
log_kind	string	Log resolution: qlog (10 Hz) or rlog (100 Hz)
log_hz	int	CAN signal sampling rate
vid_kind	string	Camera source type
camera_fps	int	Video frame rate (20 fps)
clip_id	int	Clip index within route (0-indexed)
event_mono	int	Monotonic timestamp of takeover (ns)
video_time_s	float	Takeover time within full route video (s)
clip_start_s	float	Clip start time within route (s)
clip_dur_s	float	Clip duration (s)
seg_nums_used	list	Route segments the clip was cut from

🚗 carState.csv — Vehicle Dynamics & Driver Inputs

Column	Unit	Description
vEgo	m/s	Ego vehicle speed
aEgo	m/s²	Ego vehicle acceleration
steeringAngleDeg	deg	Steering wheel angle
steeringTorque	N·m	Driver steering torque
steeringPressed	bool	Driver actively steering
gasPressed	bool	Gas pedal pressed
brakePressed	bool	Brake pedal pressed
cruiseState.enabled	bool	Cruise / ADAS engaged

🤖 controlsState.csv — ADAS Controller

Column	Unit	Description
enabled	bool	ADAS system enabled
active	bool	ADAS actively controlling vehicle
curvature	1/m	Current path curvature
desiredCurvature	1/m	Target curvature from planner
vCruise	m/s	Set cruise speed
longControlState	enum	Longitudinal control state
alertText1	string	Primary driver alert
alertText2	string	Secondary driver alert

🎮 carControl.csv — Control Commands

Column	Unit	Description
latActive	bool	Lateral control active
longActive	bool	Longitudinal control active
actuators.accel	m/s²	Commanded acceleration
actuators.torque	N·m	Commanded steering torque
actuators.curvature	1/m	Commanded path curvature

⚙️ carOutput.csv — Actuator Outputs

Column	Description
actuatorsOutput.accel	Acceleration actuator output
actuatorsOutput.brake	Brake actuator output
actuatorsOutput.gas	Gas actuator output
actuatorsOutput.steer	Steering actuator output
actuatorsOutput.steerOutputCan	Raw CAN steering output
actuatorsOutput.steeringAngleDeg	Steering angle output (deg)

🧠 drivingModelData.csv — Driving Model Predictions

Column	Description
action.desiredCurvature	Model-predicted desired curvature
action.desiredAcceleration	Model-predicted desired acceleration
laneLineMeta.leftProb	Left lane line detection probability
laneLineMeta.rightProb	Right lane line detection probability

📡 radarState.csv — Lead Vehicle Detection

Column	Unit	Description
leadOne.dRel	m	Distance to primary lead vehicle
leadOne.vRel	m/s	Relative velocity of lead
leadOne.vLead	m/s	Absolute velocity of lead
leadOne.aLeadK	m/s²	Lead vehicle acceleration
leadTwo.*	—	Secondary lead vehicle (same fields)

📐 accelerometer.csv — IMU Data

Column	Unit	Description
acceleration.v	m/s²	3-axis acceleration vector
timestamp	—	Sensor timestamp

📏 longitudinalPlan.csv — Planner Outputs

Column	Unit	Description
aTarget	m/s²	Target acceleration
hasLead	bool	Lead vehicle detected
fcw	bool	Forward collision warning active
speeds[]	m/s	Planned speed profile
accels[]	m/s²	Planned acceleration profile

🧭 VehicleIMU.csv, 🌀 Gyroscope.csv, 📷 CameraOdometry.csv, 🎯 LiveCalibration.csv, 🔧 LiveParameters.csv

Body-frame IMU and yaw rate, raw angular rates, visual odometry, the device-to-vehicle calibration, and the online vehicle-parameter estimates of the openpilot stack, each with the same time base as the other CSV files. Column names follow the corresponding openpilot cereal message fields.

---

🚀 Quick Start

Loading a Single Clip

import json
import pandas as pd
from huggingface_hub import hf_hub_download

repo_id = "HenryYHW/ADAS-TO-Sample"

# pick a clip from the annotation table
labels = pd.read_csv(hf_hub_download(repo_id, "annotations/clip_final_labels.csv", repo_type="dataset"))
clip_path = labels.loc[labels.final_label == "cover", "clip_path"].iloc[0]

# 📋 metadata
with open(hf_hub_download(repo_id, f"{clip_path}/meta.json", repo_type="dataset")) as f:
    meta = json.load(f)

# 🚗 vehicle state signals
car_state = pd.read_csv(hf_hub_download(repo_id, f"{clip_path}/carState.csv", repo_type="dataset"))
print(meta["car_model"], car_state[["vEgo", "aEgo", "steeringAngleDeg", "brakePressed"]].describe())

💾 Download the Whole Subset

# Using huggingface-cli (recommended)
huggingface-cli download HenryYHW/ADAS-TO-Sample --repo-type dataset --local-dir ./ADAS-TO-Sample

# Using git-lfs
git lfs install
git clone https://huggingface.co/datasets/HenryYHW/ADAS-TO-Sample

---

📦 Full Dataset

The full ADAS-TO corpus contains 16,446 takeover clips from 364 drivers, 179 vehicle models and 2,585 routes (~41 GB), plus the safety-critical case annotations.

👉 Access the full dataset: HenryYHW/ADAS-TO

For questions or full dataset access, contact: henryyuhangwang@gmail.com

---

🔒 Privacy & Ethics

• Anonymized identifiers: driver and route IDs are replaced with anonymous tokens (driver_NNN, route_MMM); the mapping to real device IDs is not released.
• Forward-view only: video captures the road-facing view only — no cabin or driver footage.
• No GPS: clip signal files contain no location coordinates.
• Other road users may appear in the forward video; use accordingly.

---

📝 Citation

If you use ADAS-TO in your research, please cite the arXiv paper (arXiv:2603.06986):

@article{wang2026adasto,
  title   = {ADAS-TO: A Large-Scale Multimodal Naturalistic Dataset and
             Empirical Characterization of Human Takeovers during ADAS Engagement},
  author  = {Wang, Yuhang and Xu, Yiyao and Sun, Jingran and Zhou, Hao},
  journal = {arXiv preprint arXiv:2603.06986},
  year    = {2026},
  url     = {https://arxiv.org/abs/2603.06986}
}

---

📄 License

This dataset is released under CC BY-NC 4.0.

For academic and non-commercial research purposes.

---

Built with ❤️ for the autonomous driving research community