PersonaDrive: Style-Instructed Driving Dataset

Mahmoud Srewa · Praneetsai Vasu Iddamsetty · Mohammad Abdullah Al Faruque · Salma Elmalaki

University of California, Irvine · WACV 2027

Project page · Paper (arXiv 2606.12616) · Model · Code

This dataset contains driving recorded in CARLA under three explicit style instructions: normal (neutral), aggressive and conservative. It forms the per-style retrieval databases and the style fine-tuning data of PersonaDrive.

PersonaDrive retrieves frames of the selected style as in-context demonstrations, so switching style only means searching a different part of this dataset.

Style names. The folders use the code's names:

• default is normal, everyday driving (the neutral instruction);
• fast is aggressive;
• slow is conservative.

The code also has a retrieval profile called normal. It is not part of this dataset: it is the style-agnostic index built from SimLingo's PDM-lite data.

At a glance

Folder	Style	Frames
default	Neutral: normal driving	9,613
fast	Aggressive	7,143
slow	Conservative	13,021
Total		29,777

These are the frames the 15 participants drove (driver1–driver15). The fast and slow folders also contain lb1_split: routes of SimLingo's Leaderboard 1.0 split, recorded in the aggressive and conservative style and used for Stage-3 fine-tuning. They are not part of the participant data and are not counted above.

Each frame has a front camera image, the ego state and controls, the navigation targets and the surrounding objects. Most routes also have per-frame commentary and VQA annotations.

Files

Every zip unpacks into the same directory tree, so extract the zips you need into one data root.

File	Size	Files	Contents
data_default.zip	2.37 GB	19,394	Normal (neutral) driving: camera frames, per-frame measurements, surrounding objects.
data_fast.zip	11.69 GB	81,033	Aggressive driving (fast/driver*), plus the fast/lb1_split fine-tuning routes with their retrieved context points (icl_examples.json.gz).
data_slow.zip	18.84 GB	130,004	Conservative driving (slow/driver*), plus the slow/lb1_split fine-tuning routes, organized like data_fast.zip.
commentary.zip	3.2 MB	786	Per-route commentary.json for all three styles.
drivelm.zip	27.8 MB	775	Per-route vqa.json (DriveLM-style question–answer pairs) for all three styles.

Total: 32.94 GB. SHA256SUMS lists the checksum of each zip.

# everything
hf download PervasiveAutonomyLab/PersonaDrive --repo-type dataset --local-dir personadrive_data
# or only some parts, e.g. normal driving with its language annotations
hf download PervasiveAutonomyLab/PersonaDrive data_default.zip commentary.zip drivelm.zip SHA256SUMS \
    --repo-type dataset --local-dir personadrive_data

cd personadrive_data && sha256sum -c SHA256SUMS --ignore-missing
for z in *.zip; do unzip -q "$z" -d "$PERSONADRIVE_DATA_ROOT"; done

Which zips you need:

Task	Zips
Use the driving data of one style	data_<style> (plus commentary / drivelm for language)
Evaluate PersonaDrive with a style (default / fast / slow)	data_<style> of that style
Re-run Stage-3 fine-tuning	data_fast, data_slow, commentary, drivelm, plus SimLingo's PDM-lite data
Re-run Stages 1–2 (retrieval head, per-style indices)	all five zips, plus PDM-lite for the style-agnostic normal index

Evaluation with retrieval opens the camera frames of the retrieved demonstrations, so it needs the frames of the chosen style. It also needs the retrieval head, which is in the code repository, and the per-style FAISS indices, which you build from this dataset (Build the retrieval indices). The code repository's datasets/README.md lists which zips each use needs.

How the data was collected

15 participants drove CARLA Leaderboard routes at a driver-in-the-loop rig with a real wheel and pedals. The recording code is released with PersonaDrive (human_data_collection/). They drove the same routes three times, each time under a different instruction:

Folder	Style	What the driver was told
default	Neutral: normal driving	"Follow traffic laws, navigate safely, and drive at a reasonable speed."
fast	Aggressive	"Drive as fast as possible without crashing; engage in aggressive behaviors such as tailgating and cutting off merging vehicles."
slow	Conservative	"Target 10–15 mph on surface streets and 25–35 mph on the highway; yield to other vehicles whenever possible."

Rig.

• Hardware: a Next Level Racing GTrack bucket seat, three 50-inch monitors covering about 130° of view, and a Logitech G923 wheel and pedals.
• Simulator: CARLA 0.9.15 ran synchronously at 20 Hz. The ego vehicle was the only human-controlled actor, and CARLA's Traffic Manager populated the surrounding traffic.
• Display: a Pygame window showed the first-person view, the speed, the active navigation instruction and scenario, and a bird's-eye view of the surrounding vehicles (for manoeuvres such as merges).

Protocol.

• Disjoint from evaluation: routes were drawn at random from the CARLA Leaderboard route sets, and none of them is a Bench2Drive evaluation route. They share the scenario taxonomy, so a retrieved demonstration can be relevant to an evaluation situation, but no evaluation episode is in the dataset.
• Three passes: each participant drove the routes once per instruction, in randomized order, so that style is not confounded with route familiarity.
• Supervision: participants were supervised to make sure they followed the instructions.
• Anonymity: driver<N> is an anonymous participant index. The same index refers to the same participant in all three styles.

Style	Mean speed (mph)	In motion (mph)	Std. dev.	Stopped (%)
Conservative (slow)	9.8	11.5	7.6	14.6
Neutral, normal driving (default)	13.2	15.6	10.8	15.7
Aggressive (fast)	17.5	21.8	16.8	19.8

_{Speed statistics computed from the recorded ego state, route-matched over the 55 routes driven under all three instructions. The instructions separate before any model is involved.}

What is recorded. Each route stores, at 4 Hz:

• the front camera image, plus a shifted-camera copy used for augmentation;
• the ego state from CARLA (speed, pose), which follows the participant's driving;
• the controls throttle, steer and brake in measurements.json. These are the controls of SimLingo's privileged autopilot, which runs alongside the participant on every frame. The participant's own wheel and pedal inputs are recorded by the rig but are not part of this release;
• the navigation command and the next GPS target points from CARLA's route planner;
• the objects around the ego vehicle: vehicles, pedestrians, traffic signs, traffic-light states.

commentary.zip: what the car is doing, in words

• Contents: one sentence per frame that says what the car does and why, together with the object that causes it. For example: "Follow the route. Accelerate to drive with the target speed."
• How it was made: SimLingo's commentary generator produced it automatically after recording, from the recorded simulator state (route, hazards, surrounding objects). Participants did not write it.

drivelm.zip: questions and answers about each scene

• Contents: per-frame question–answer pairs about perception, prediction and planning, in the DriveLM format. For example: Q: "What are the important objects in the scene?" A: "The important object in the scene is the black car to the front left of the ego vehicle."
• How it was made: a DriveLM-style generator produced them automatically from the recorded objects and ego state.

Layout

simlingo/
├── data/simlingo/
│   ├── default/driver<N>/routes/<Scenario>/<id>/<Town..>/            # normal (neutral) driving
│   ├── fast/driver<N>/routes/<Scenario>/<id>/<Town..>/               # aggressive
│   ├── fast/lb1_split/routes_training/<Scenario>/<route>/<Town..>/   # SimLingo lb1_split routes, aggressive (fine-tuning)
│   ├── slow/driver<N>/routes/<Scenario>/<id>/<Town..>/               # conservative
│   └── slow/lb1_split/routes_training/<Scenario>/<route>/<Town..>/   # SimLingo lb1_split routes, conservative (fine-tuning)
├── commentary/simlingo/<same relative paths>/commentary.json
└── drivelm/simlingo/<same relative paths>/vqa.json

Each route folder (<Town..>/) contains:

File	Contents
rgb/NNNN.jpg	Front camera, one frame per dataset step (4 Hz).
rgb_augmented/NNNN.jpg	Shifted-camera augmentation.
measurements.json	Per-frame list: speed, throttle, steer, brake, command, next_command, target_point(_next), ego_matrix, route, aim_wp, …
boxes.json	Per-frame surrounding objects (driver<N> routes).
result.json, records.json.gz	Simulator logs.
icl_examples.json.gz	lb1_split routes only. Written by Stage 3a of the code: for every frame, the K = 2 most similar frames retrieved from the same style's database, i.e. the demonstrations the model was fine-tuned with (rgb_path, speed_history, throttle_history, command, target_point(_next), commentary, waypoints).

commentary.json and vqa.json are per-frame lists aligned with measurements.json; frames without an annotation are null.

Processing and privacy

• Anonymity: participants are identified only by an index (driver1–driver15). The dataset contains no names or other personal information about participants.
• Frame paths in annotations: the frame paths inside commentary.json and vqa.json were rewritten to paths relative to the data root (simlingo/data/simlingo/...).
• Paths in icl_examples.json.gz: these keep the absolute paths of the machine they were built on. The PersonaDrive loader resolves both forms against $PERSONADRIVE_DATA_ROOT.

Not included

• SimLingo's PDM-lite data (training_1_scenario for Stage-3 fine-tuning; training_1_scenario, training_3_scenarios and SimLingo's own lb1_split for the normal index, see the code repository's datasets/README.md). It is the SimLingo dataset of Wayve Technologies Ltd and is distributed by its authors under their own licence; download it from SimLingo.
• The FAISS indices and other retrieval artefacts. Build them from this dataset with the code repository.

Citation

@article{srewa2026personadrive,
  title   = {PersonaDrive: Human-Style Retrieval-Augmented VLA Agents for Closed-Loop Driving Simulation},
  author  = {Srewa, Mahmoud and Iddamsetty, Praneetsai Vasu and Al Faruque, Mohammad Abdullah and Elmalaki, Salma},
  journal = {arXiv preprint arXiv:2606.12616},
  year    = {2026}
}

Acknowledgements

This work is supported by the U.S. National Science Foundation under grant 2339266 and partially supported by the UCI ProperAI Institute.