immanuelpeter/carla-autopilot-multimodal-dataset

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}
}