🚗 BATON

A Multimodal Benchmark for Bidirectional Automation Transition Observation in Naturalistic Driving

     

A large-scale, real-world multimodal dataset and benchmark for predicting when drivers hand control to driving automation and when they take it back.
Under review at the KDD 2027 Datasets and Benchmarks Track.

---

_{(a) In-vehicle collection setup · (b) synchronized road video, cabin video, CAN signals and GPS around one takeover and one handover · (c) global route distribution · (d) the three benchmark tasks.}

---

🎬 Live Preview

Continuous sequence — cabin fisheye · front view · 5 fps

8 CAN/IMU sensor streams cycling through all channels

Daytime time-lapse — front · cabin · 2-min intervals

_{Nighttime driving — time-lapse with ⬆ DAS Handover and ↩ Human Takeover event highlights}

---

📊 Dataset at a Glance

Released corpus

🌍 Routes	👤 Drivers	🚙 Car Models	⏱️ Driving	🤖 DAS-engaged	🧑 Human-driven	🗺️ Coverage
781	173	108	204.9 h	49.0%	51.0%	5 continents

Frozen benchmark subset (all reported results; fixed across dataset versions)

📦 Route bundles	👤 Drivers	🚙 Car Models	⏱️ Driving	🔄 Transitions	⬆️ Handovers	↩️ Takeovers
565	150	99	162.1 h	3,593	1,800	1,793

_{Global distribution of routes, per-driver driving duration, and the control-transition breakdown of the benchmark subset.}

BATON is a living dataset: collection continues, and additions are released under the same two-tier process as version-tagged, additive updates. The benchmark subset, its splits and all reported results stay frozen so that results remain comparable across versions.

What this repository hosts. The gated raw tier: per-driver folders with front and cabin video (qcamera.mp4, dcamera.mp4), per-route sensor CSVs and metadata, plus the benchmark/ sample-definition files. Access is granted after identity verification. A 43-route sample with all modalities is available without application at BATON-Sample; code, splits and benchmark files are on GitHub.

---

🔬 Data Collection & Modalities

Setup: Non-intrusive plug-and-play comma device (OBD-II CAN access + dual cameras). Drivers use their own vehicles during real daily driving — no lab, no script.

Component	Spec
📡 CAN bus	decoded vehicle, controller and planner signals at up to 100 Hz
📷 Front camera	526×330 · H.264 · 20 fps
🎥 Cabin fisheye	1928×1208 · HEVC · 20 fps
🛰️ GPS	10 Hz, with derived route context (road type, speed limit, lane count, intersection/ramp proximity)

Synchronized modalities per route:

• vehicle_dynamics.csv — speed, acceleration, steering, pedals, DAS engagement state
• planning.csv — planner curvature, lane-change state, lead flag
• radar.csv — lead-vehicle distance and relative speed
• driver_state.csv — derived driver-monitoring outputs (face pose, eye state, awareness)
• imu.csv — 3-axis accelerometer and gyroscope
• gps.csv, localization.csv — position, heading, road context
• qcamera.mp4 — front-view video
• dcamera.mp4 — in-cabin fisheye video

📷 Front · Day	📷 Front · Night	⬆️ DAS Handover
🎥 Cabin · Day	🎥 Cabin · Night	↩️ Human Takeover

_{Multimodal context around a transition: map with transition markers, cabin and road frames, and the aligned signal tracks (vehicle dynamics, driver inputs, radar, driver monitoring, planner) in the 30 s around one takeover.}

---

🏆 Benchmark Tasks

_{(a) Task-1 action classes · (b, c) positive and negative windows for the handover and takeover tasks.}

Task	Description	Windows (h = 3 s)	Labels	Primary metric
🎯 Task 1	Signal-derived driving-action recognition (auxiliary; official protocol uses rule-free inputs; a multi-label variant is released)	1,161,794	Cruising · Car following · Accelerating · Braking · Lane change · Turning · Stopped	Macro-F1
⬆️ Task 2	Handover prediction (human → DAS)	65,223	Handover (15.8%) · No handover	AUPRC (sample / event)
↩️ Task 3-D	Takeover onset detection (DAS → human), leak-safe inputs	91,255	Takeover (12.0%) · No takeover	AUPRC (sample / event)
↩️ Task 3-A	Takeover pre-override anticipation: driver-override channels withheld, windows end ≥ 1 s before the override onset	subset of Task 3	Takeover (3.5%) · No takeover	AUPRC (sample / event)

Leak-safe protocol. Handover and takeover labels are defined from the automation-engagement state, so four ADAS-control variables (cc_latActive, cruiseState_enabled, cs_longControlState, actuators_accel) are withheld from all inputs. Task 3-A additionally withholds the six driver-override channels (brake / gas / steering pressed flags, steering torque, brake, gas), separating genuine anticipation from recognition of an override already in progress.

Reference results (cross-driver, h = 3 s, sample AUPRC, 3 seeds). Handover: 0.335 (BATON-WM: CAN statistics + frozen video world-model features) vs 0.236 for the tabular baseline. Takeover onset detection: 0.514 (CAN statistics + pose). Takeover anticipation: 0.070 against a 0.035 base rate — genuine pre-override anticipation remains an open problem.

---

📐 Evaluation Protocol

Setting	Value
Primary split	Cross-driver: disjoint drivers in train / val / test (405 / 84 / 76 routes; 104 / 26 / 20 drivers). Because each driver drives their own vehicle, this is a driver–vehicle pair hold-out.
Additional splits	Cross-vehicle (vehicle models held out), within-driver temporal, random
Input window	5 s, stride 0.5 s
Prediction horizon	1 s, 3 s, 5 s (main: 3 s)
Random seeds	42, 123, 7 — 3-seed mean ± std
Metrics	Task 1: Macro-F1 (and macro-AP for the multi-label variant). Tasks 2 / 3: sample- and event-level AUPRC, recall at false-alarm budgets, warning lead time

---

📁 Code and Benchmark Files (GitHub)

BATON/                              https://github.com/OpenLKA/BATON
├── benchmark_v2/                   # frozen benchmark: routes, labels, task CSVs, splits, checksums
│   ├── task1_action_samples.csv, task2_activation_samples_h{1,3,5}.csv, task3_takeover_samples_h{1,3,5}.csv
│   ├── task3_takeover_samples_h3_antsafe*.csv      # Task 3-A (anticipation-safe) definitions
│   ├── split_cross_driver.json, split_cross_vehicle.json, split_within_device_temporal.json, split_random.json
│   └── CHECKSUMS.sha256, benchmark_protocol.md
├── benchmark/                      # generation pipeline and statistics scripts
├── baseline/                       # XGBoost, GRU/TCN, V-JEPA2 fusion, VLM baselines, multi-label Task 1
├── data_processing/                # feature extraction (V-JEPA2, pose, GPS road context)
├── REPRODUCE.md · DATASHEET.md · LICENSE

---

🚀 Quick Start

# Sample dataset (43 routes, all modalities, no application needed)
git lfs install
git clone https://huggingface.co/datasets/HenryYHW/BATON-Sample

# Full raw tier (this repository; gated — request access on the dataset page first)
python -c "
from huggingface_hub import snapshot_download
snapshot_download('HenryYHW/BATON', repo_type='dataset', local_dir='./data')
"

# Code, benchmark definitions and reproduction steps
git clone https://github.com/OpenLKA/BATON && cat BATON/REPRODUCE.md

---

📡 Data Access

Resource	Link
📦 Full raw tier (gated)	HuggingFace — HenryYHW/BATON
🔍 Sample dataset (43 routes)	HuggingFace — HenryYHW/BATON-Sample
💻 Code and benchmark files	GitHub — OpenLKA/BATON
📄 Paper	arxiv.org/abs/2604.07263

Privacy. Faces of non-driver occupants and roadside faces and plates are blurred; raw GPS traces are released only as derived route context. Contributors can withdraw their data at any time, and withdrawals propagate to every tier. Please do not attempt to re-identify drivers or use the data for surveillance.

---

📜 Citation

@article{wang2026baton,
  title   = {BATON: A Multimodal Benchmark for Bidirectional Automation Transition
             Observation in Naturalistic Driving},
  author  = {Wang, Yuhang and Xu, Yiyao and Yang, Chaoyun and Li, Lingyao
             and Sun, Jingran and Zhou, Hao},
  journal = {arXiv preprint arXiv:2604.07263},
  year    = {2026}
}

---

📄 License

The data are released for academic research use only under CC BY-NC 4.0; the code is MIT-licensed (see the GitHub repository).

• Attribution — cite the BATON paper (see Citation above) in any publication or work that uses this dataset.
• Non-commercial — commercial use of this dataset or any derivative is not permitted.
• No re-identification — do not attempt to identify drivers or use the data for surveillance.

For commercial licensing inquiries, please contact the authors.

---

_{🔗 Paper  ·  Full Dataset  ·  Sample Dataset  ·  GitHub}