metadata

dataset_info:
  features:
    - name: run_id
      dtype: string
    - name: frame
      dtype: int32
    - name: timestamp
      dtype: float32
    - name: image_front
      dtype: image
    - name: image_front_left
      dtype: image
    - name: image_front_right
      dtype: image
    - name: image_rear
      dtype: image
    - name: seg_front
      dtype: image
    - name: lidar
      list:
        list: float32
    - name: boxes
      list:
        list: float32
    - name: box_labels
      list: string
    - name: location_x
      dtype: float32
    - name: location_y
      dtype: float32
    - name: location_z
      dtype: float32
    - name: rotation_pitch
      dtype: float32
    - name: rotation_yaw
      dtype: float32
    - name: rotation_roll
      dtype: float32
    - name: velocity_x
      dtype: float32
    - name: velocity_y
      dtype: float32
    - name: velocity_z
      dtype: float32
    - name: speed_kmh
      dtype: float32
    - name: throttle
      dtype: float32
    - name: steer
      dtype: float32
    - name: brake
      dtype: float32
    - name: nearby_vehicles_50m
      dtype: int32
    - name: total_npc_vehicles
      dtype: int32
    - name: total_npc_walkers
      dtype: int32
    - name: map_name
      dtype: string
    - name: weather_cloudiness
      dtype: float32
    - name: weather_precipitation
      dtype: float32
    - name: weather_fog_density
      dtype: float32
    - name: weather_sun_altitude
      dtype: float32
    - name: vehicles_spawned
      dtype: int32
    - name: walkers_spawned
      dtype: int32
    - name: duration_seconds
      dtype: int32
  splits:
    - name: train
      num_bytes: 298274262201
      num_examples: 67000
    - name: validation
      num_bytes: 35503432435.4
      num_examples: 8400
    - name: test
      num_bytes: 31770625008.6
      num_examples: 7200
  download_size: 361766155632
  dataset_size: 365548319645
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: validation
        path: data/validation-*
      - split: test
        path: data/test-*
license: mit
task_categories:
  - object-detection
  - image-classification
  - image-segmentation
  - depth-estimation
  - video-classification
  - any-to-any
  - image-to-text
  - reinforcement-learning
language:
  - en
pretty_name: CARLA Autopilot Multimodal Dataset
size_categories:
  - 10K<n<100K

CARLA Autopilot Multimodal Dataset

This dataset contains synchronized multimodal driving data collected in the CARLA simulator using the autopilot feature. It provides RGB images from multiple cameras, semantic segmentation, LiDAR point clouds, 2D bounding boxes, and ego-vehicle state/control signals across varied weather, maps, and traffic densities.

The dataset is designed for research in autonomous driving, sensor fusion, imitation learning, and self-driving evaluation.

Dataset Summary

Runs: 30 autopilot runs
Sensors:
- RGB cameras: front, front-left, front-right, rear (800×600, fov=90°)
- Semantic segmentation: front (raw + colorized)
- LiDAR: 32-channel ray-cast, 20 Hz, 80 m range
- Collision sensor for impact logs
Annotations: 2D bounding boxes and class labels (vehicles, pedestrians) w.r.t front camera
Ego states: position, rotation, velocity, control (throttle/steer/brake), speed (km/h)
Environment: varied weather, time-of-day (sun altitude), NPC traffic (vehicles + pedestrians)

Splits

Train: 67,000 frames
Validation: 8,400 frames
Test: 7,200 frames
Total size: ~365 GB

Relation to Previous Versions

This dataset, CARLA Autopilot Multimodal Dataset, is an extension of the earlier CARLA Autopilot Image Dataset.

Previous version (carla-autopilot-images):
Contained synchronized RGB camera views (front, front-left, front-right, rear) with ego-vehicle states, controls, and environment metadata.
Current version (carla-autopilot-multimodal-dataset):
Adds new sensor modalities and richer annotations, including:
- Semantic segmentation (front view)
- LiDAR point clouds
- 2D bounding boxes and labels (vehicles, pedestrians)
- Expanded metadata (collisions, weather difficulty, quality metrics)

In short, v2 augments the original dataset with multimodal signals for perception + sensor fusion research, while retaining full compatibility with the core camera + state data from v1.

Features

Each sample contains:

run_id (string): Identifier for the simulation run
frame (int): Frame number
timestamp (float): Relative timestamp (s)
image_front, image_front_left, image_front_right, image_rear (images): RGB views
seg_front (image): Semantic segmentation (front view)
lidar (list[list[float32]]): LiDAR point cloud (x, y, z, intensity)
boxes (list[list[float32]]): 2D bounding boxes in [xmin, ymin, xmax, ymax] format
box_labels (list[string]): Class labels for bounding boxes
location_{x,y,z} (float): Ego position in world coords
rotation_{pitch,yaw,roll} (float): Ego rotation
velocity_{x,y,z} (float): Ego velocity (m/s)
speed_kmh (float): Ego speed (km/h)
throttle, steer, brake (float): Control inputs
nearby_vehicles_50m, total_npc_vehicles, total_npc_walkers (int): Traffic counts
map_name (string): CARLA map used
weather_* (float): Weather conditions (cloudiness, precipitation, fog, sun altitude)
vehicles_spawned, walkers_spawned (int): Number of NPCs
duration_seconds (int): Total run length in seconds

Example Usage

from datasets import load_dataset

ds = load_dataset("immanuelpeter/carla-autopilot-multimodal-dataset", split="train")
sample = ds[0]

# Access RGB image and LiDAR
front_img = sample["image_front"]
lidar = sample["lidar"]
boxes = sample["boxes"]

Collection Process

Data was collected using a custom CARLA Python script that:

Spawns an ego vehicle with autopilot enabled
Spawns configurable NPC vehicles and pedestrians
Randomizes weather and lighting conditions per run
Synchronizes all sensors and saves every N-th frame
Records vehicle state, control signals, collisions, and environment statistics

All sensors operate in synchronous mode for frame alignment.

Intended Use

Training and benchmarking multimodal self-driving models
Research on sensor fusion, perception, and planning
Imitation learning from autopilot trajectories
Evaluation under diverse weather and traffic conditions

Citation

If you use this dataset, please cite the CARLA simulator:

@inproceedings{Dosovitskiy17,
  title = {CARLA: An Open Urban Driving Simulator},
  author = {Alexey Dosovitskiy and German Ros and Felipe Codevilla and Antonio Lopez and Vladlen Koltun},
  booktitle = {Proceedings of the 1st Annual Conference on Robot Learning},
  pages = {1--16},
  year = {2017}
}