L2D (Learning to Drive): Multimodal Self-Driving Dataset
- Provider
- Yaak and Hugging Face LeRobot
- License
- Apache-2.0 (source)
- Size
- 9,651 files
- Format
- Parquet (in sampled files)
- Stored at
- Hugging Face (https://huggingface.co/datasets/yaak-ai/L2D)
- Access
- Free and public. No account with the data provider is needed.
- Category
- Robotics
- Source
- https://huggingface.co/blog/lerobot-goes-to-driving-school
About L2D (Learning to Drive): Multimodal Self-Driving Dataset
L2D (Learning to Drive) is a large multimodal driving dataset released by Yaak with the Hugging Face LeRobot team for training end-to-end driving policies. Data come from driving-school cars (KIA Niro EV 2023) in 30 German cities, with expert policies from driving instructors and student policies from learner drivers. Each episode carries six surround-view camera streams plus a rendered map view, vehicle state (speed, heading, GPS, IMU), continuous actions (gas, brake, steering) and discrete actions (gear, turn signals), environment state such as lane count, road type, road surface and speed limit, weather and lighting conditions, future waypoints snapped to OpenStreetMap, and a natural-language instruction per episode. The repository is a LeRobot v3.0 dataset: meta/info.json lists 100,000 episodes, 26.5 million frames at 10 fps and 4,219 distinct tasks; tabular data are Parquet under data/ and camera streams are MP4 under videos/. Measured 2026-11-14, the repo holds about 5.6 TB. Open it with lerobot's LeRobotDataset, or read the Parquet files with pandas or pyarrow. Useful for imitation learning, vision-language-action models, world models and language-conditioned driving. The card states the licence as Apache-2.0; cite the Yaak and LeRobot release.
How to load L2D (Learning to Drive): Multimodal Self-Driving Dataset in Python
from datasets import load_dataset
ds = load_dataset("yaak-ai/L2D")
print(ds)Data structure
File types in sampled files: Parquet
Files
Example file: data/chunk-000/file-000.parquet
Analyze this dataset without downloading it
On Scigantic, L2D (Learning to Drive): Multimodal Self-Driving Dataset opens in a hosted JupyterLab notebook with the data already mounted, so you can read files where they are stored instead of copying them first. Each notebook includes an AI assistant that can see the mounted files and write and run analysis code. Open the dataset page to launch a notebook, or browse the full catalog.
This page is also available as Markdown and as JSON.
Related Robotics datasets
- Eidon Tracker POV: Egocentric Household Video with IMU Arm Tracking
- PhysicalAI Robotics Manipulation Objects Kitchen MJCF
- IsaacLab-Arena Unitree G1 Humanoid Apple Pick-and-Place: 200 XR-Teleoperated Simulation Demonstrations (NVIDIA)
- GrandTour: Legged Robot Multi-Sensor Dataset (ETH Zurich)
- PhysicalAI Traffic Anomaly Reasoning
- PhysicalAI Robotics Manipulation Kitchen Demos
- Dexora Real-World Dataset: High-DoF Bimanual Dexterous Teleoperation
- PhysicalAI Robotics GR 00 T GR 1