DriveDNA — A Large-Scale Benchmark for Personalized Driving Style

DriveDNA turns a large, in-the-wild naturalistic driving corpus into the first benchmark for personalized driving style: representing who is driving (the driver-specific, vehicle- and context-invariant signature) as distinct from what they are driving and where. It pairs time-synchronized CAN telemetry and forward-road video across hundreds of drivers and vehicle models, and — uniquely — cleanly separates human-driven from ADAS-engaged driving.

Built on the openpilot/comma corpus at ../Dataset (see that folder's README.md for the raw data dictionary). This document introduces the benchmark derived from it.

---

Highlights

Drivers (unique dongle IDs)	465
Vehicle models	120 consolidated (from 218 raw fingerprint variants), 26 brands
Routes	4,373 continuous drives (~1-min segments)
Telemetry / video	~1,938 h CAN + ~2,027 h forward camera
Human-vs-ADAS	per-frame control-authority labels (cs_enabled / cruiseState_enabled)
Style-usable subset	614 h human-driven & moving, 460 drivers (10 Hz decoded)
Curated benchmark	580 h · 355 drivers · 24 within-vehicle models with frozen official splits

Why it's unique. Every public personalized-style resource is tiny and holds vehicle/route fixed (PDB 2025 = 12 drivers/1 car; UAH = 6; HCRL/KCID CAN-ID sets ≤16); large AV datasets (nuScenes, Waymo, nuPlan) carry no driver identity. DriveDNA is the first public corpus combining many drivers × many vehicles × multi-session CAN+video, with human-vs-ADAS separation no other dataset offers.

---

Modalities & committed signals

All streams are decoded from the openpilot logs and resampled to a unified 10 Hz grid. The benchmark commits to a signal set chosen for driving-style research (each maps to an established construct):

Signal	Meaning	Style construct
vEgo, aEgo (+ jerk)	speed, longitudinal accel	longitudinal aggressiveness (jerk, RMS accel)
steeringAngleDeg, steeringPressed	driver steering INPUT (vehicle-dependent via steerRatio)	steering entropy, reversal rate
actual_curvature	realized path curvature (vehicle-normalized trajectory)	cornering sharpness, path geometry
yaw_rate → curv_measured, slip	independently-sensed turning; steering-vs-realized slip	limit/aggressiveness
leadOne_dRel/vLead/vRel	lead-vehicle distance & relative speed (radar)	time-headway (THW), TTC, gap preference
gas, brake (+ pressed)	pedal application	pedal dynamics
laneLeft_y, laneRight_y	lane offsets	lane-keeping (SDLP), lane-position bias
cs_enabled, cruiseState_enabled	openpilot / OEM-ACC engagement	human-vs-ADAS gate

Key distinction — steering INPUT vs realized PATH. steeringAngleDeg is the raw wheel input and is vehicle-dependent (a high-steerRatio car needs more angle for the same path). actual_curvature (controlsState.curvature) is the vehicle-normalized realized path. Their map is the vehicle transfer function — a signal-level handle on the "who vs which-car" question at the heart of the benchmark.

Human-vs-ADAS. Style is learned only from human-driven frames (cs_enabled==0 AND cruiseState_enabled==0); ADAS-engaged frames (≈53 % of driving in this ADAS-heavy population) are kept for a separate human↔automation analysis.

---

Benchmark tasks & splits

Frozen official splits live in data/splits/ (deterministic, seed-locked). Vehicle grouping uses the consolidated model names (code/preprocessing/model_merge.py), so year/trim variants of one nameplate count as the same vehicle.

• Within-vehicle driver identification / verification — hold the vehicle model fixed (24 eligible consolidated models, ≥4 multi-trip drivers each) so any driver signal must be style, not vehicle. Route-grouped, held-out-trip evaluation. The leakage-free identity test.
• Cross-route / cross-session — disjoint trips/days of the same drivers (179 drivers). Removes route leakage.
• Cross-vehicle transfer — drivers who span ≥2 consolidated models (16 drivers). Same-driver style across different cars (secondary, honest case study).
• Human ↔ ADAS — manual vs ADAS-engaged partitions.
• Personalized behavior prediction / few-shot adaptation — driver-disjoint folds (train 212 / val 45 / test 45) + 53 held-out drivers for few-shot (infer a driver's style from a few minutes, then personalize prediction).

---

Validated baseline findings (handcrafted features)

Cheap handcrafted-feature baselines already establish the benchmark's premise on the clean corpus (learned models forthcoming):

• Style is identifiable within a vehicle. Route-grouped driver-ID reaches 8.0× chance on held-out trips (RIVIAN R1, 14 drivers) — and survives removing all speed features, so it is genuine control style.
• The leakage mechanism, at the signal level. Predicting the vehicle model: raw steering scores 2.3× chance (it encodes steerRatio) while realized actual_curvature scores ≈chance (vehicle- agnostic). This is why prior driver-ID methods leak vehicle identity and collapse cross-vehicle.
• Confound-clean. A log-source (qlog/rlog) leakage probe is at chance (balanced-acc 0.500), confirming results are not a decoding artifact.

Full numbers: experiments/logs/sanity_results.md.

---

Repository layout

DriveDNA/
├── README.md                     # this file
├── Plan.md                       # research/sprint plan
├── code/
│   ├── preprocessing/
│   │   ├── rlog_extract.py        # openpilot log → committed signals @10 Hz (qlog-preferred)
│   │   ├── decode_corpus.py       # parallel full-corpus decode → data/cache/
│   │   ├── style_features.py      # handcrafted style features (INPUT/PATH/long/follow/lane groups)
│   │   ├── availability_audit.py  # per-signal / per-model coverage
│   │   ├── curate_splits.py       # quality filters → frozen official splits
│   │   └── model_merge.py         # 218 raw fingerprints → 120 nameplate models
│   └── eval/
│       ├── identifiability.py     # within-vehicle driver-ID (route-grouped CV)
│       ├── leakage_probe.py       # steering-vs-curvature vehicle-leakage probe
│       ├── cross_vehicle_transfer.py
│       ├── log_type_probe.py      # qlog/rlog confound gate
│       └── decode_correctness.py / curvature_correctness.py
├── data/
│   ├── cache/<MODEL>/<driver>__<route>.parquet   # decoded 10 Hz signals
│   ├── splits/                    # within_vehicle / cross_route / cross_vehicle / driver_folds (JSON)
│   ├── availability_audit.csv
│   └── merged_model_stats.csv     # per-model driver/route/hour distribution (all 120)
└── experiments/logs/             # run logs + sanity_results.md

---

Quickstart (reproduce the data pipeline)

VENV=python  # openpilot venv python
export PYTHONPATH=/path/to/openpilot

# 1. decode the full corpus to 10 Hz signal parquets (uniform qlog)
$VENV code/preprocessing/decode_corpus.py --workers 6 --rate 10

# 2. per-signal availability audit  → data/availability_audit.csv
$VENV code/preprocessing/availability_audit.py

# 3. curate + freeze official splits → data/splits/
$VENV code/preprocessing/curate_splits.py

# 4. evaluate the premise
$VENV code/eval/identifiability.py RIVIAN_R1        # within-vehicle driver-ID
$VENV code/eval/leakage_probe.py RIVIAN_R1 FORD_F_150 TESLA_MODEL_3 ...   # vehicle-leakage
$VENV code/eval/log_type_probe.py                   # confound gate

Requirements: the openpilot venv (tools.lib.logreader, pycapnp), numpy/pandas/pyarrow, scikit-learn, ffmpeg.

---

Ethics & privacy

Crowdsourced dashcam data. Forward video is low-resolution (526×330), and the public release applies de-identification: face/plate blurring on cabin video, GPS coarsening, and hashed device IDs. Style is analyzed only on consenting drivers' human-driven segments. See the paper's ethics section for the consent basis and IRB status.

---

Status & citation

Data foundation is final and validated; the learned DriveDNA representation (a driver-conditioned latent world model) is in development. Citation and license will be added on release.