README.md

metadata

license: mit
task_categories:
  - robotics
  - image-to-text
tags:
  - autonomous-driving
  - carla
  - simlingo
  - behavioral-cloning
size_categories:
  - 100K<n<1M

SimLingo CARLA Dataset (Raw, 4Hz)

Raw driving data from CARLA simulator. No transformations or derived fields - all original measurements preserved as-is.

Dataset Summary

• Source: SimLingo (CVPR 2025)
• Scale: 228,757 frames (23 shards)
• Frame Rate: 4 FPS
• Resolution: 1024x512 RGB
• Routes: Complete driving episodes (routes never split across shards)

Column Schema

Core Fields

Column	Type	Description
route_id	string	Route identifier
frame_idx	int32	Frame index within route
image	bytes	Original JPEG image bytes

Control Signals (Raw)

Column	Type	Description
steer	float32	Steering [-1, 1]
throttle	float32	Throttle [0, 1]
brake	bool	Brake applied

Vehicle State

Column	Type	Description
speed	float32	Current speed (m/s)
target_speed	float32	Target speed
speed_limit	float32	Speed limit
theta	float32	Heading angle
angle	float32	Angle to target

Navigation

Column	Type	Description
command	int32	Navigation command
next_command	int32	Next navigation command
pos_global	string (JSON)	Global position [x, y]
target_point	string (JSON)	Target point
target_point_next	string (JSON)	Next target point
aim_wp	string (JSON)	Aim waypoint
route	string (JSON)	Planned route waypoints
route_original	string (JSON)	Original route waypoints
changed_route	bool	Route was changed

Hazards & Environment

Column	Type	Description
junction	bool	In junction
vehicle_hazard	bool	Vehicle hazard detected
vehicle_affecting_id	int32	ID of affecting vehicle
walker_hazard	bool	Pedestrian hazard
walker_affecting_id	int32	ID of affecting pedestrian
light_hazard	bool	Traffic light hazard
stop_sign_hazard	bool	Stop sign hazard
stop_sign_close	bool	Stop sign nearby
walker_close	bool	Pedestrian nearby
walker_close_id	int32	ID of nearby pedestrian
speed_reduced_by_obj_type	string	Object type causing speed reduction
speed_reduced_by_obj_id	int32	Object ID causing speed reduction
speed_reduced_by_obj_distance	float32	Distance to speed-reducing object
control_brake	bool	Control brake applied

Augmentation (from SimLingo)

Column	Type	Description
augmentation_translation	float32	Translation augmentation
augmentation_rotation	float32	Rotation augmentation

Transforms

Column	Type	Description
ego_matrix	string (JSON)	4x4 ego vehicle transform matrix
boxes	string (JSON)	3D bounding boxes for all objects

Commentary (Optional)

Column	Type	Description
commentary	string	Natural language commentary
commentary_data	string (JSON)	Full commentary object with metadata

Usage

from datasets import load_dataset
import json

ds = load_dataset("TESS-Computer/carla-simlingo-raw", split="train")

sample = ds[0]
print(sample['route_id'])
print(sample['steer'], sample['throttle'], sample['brake'])
print(sample['speed'])

# Parse JSON fields
pos = json.loads(sample['pos_global'])
boxes = json.loads(sample['boxes']) if sample['boxes'] else []

Data Collection

• Simulator: CARLA 0.9.15 (Leaderboard 2.0)
• Expert: PDM-Lite (rule-based, 100% route completion)
• Scenarios: Single-scenario routes with random weather
• Towns: Towns 1-13

Citation

@inproceedings{renz2025simlingo,
  title={SimLingo: Vision-Only Closed-Loop Autonomous Driving with Language-Action Alignment},
  author={Renz, Katrin and Chen, Long and Arani, Elahe and Sinavski, Oleg},
  booktitle={CVPR},
  year={2025},
}

License

MIT (dataset processing code). Original data subject to SimLingo and CARLA licenses.