L2D is a large-scale autonomous driving dataset created by yaak-ai and hosted on Hugging Face. It contains over 90 terabytes of multimodal data, representing more than 5,000 hours of driving footage collected from 30 cities in Germany. The dataset was last updated on September 29, 2025.
Use Cases
- Training end-to-end driving models based on continuous control signals (gas, brake, steering) and discrete actions (gear, turn signals).
- Developing sensor fusion algorithms based on data from six surrounding HD cameras and vehicle state information (speed, heading, GPS, IMU).
- Analyzing driving behavior and policy learning based on recorded environmental states (lane count, road type, surface, speed limit).
- Benchmarking perception systems on diverse urban and highway scenarios across multiple German cities.
Strengths
- Extremely large scale with over 90 terabytes of data and 5,000+ hours of driving.
- Multimodal data collection includes six HD cameras and complete vehicle state telemetry.
- Geographic diversity from 30 different cities in Germany.
Limitations
- Column-level documentation is absent; field semantics must be inferred after download.
- Row count, file formats, and license information are unknown, which may limit suitability assessment.
Provenance
- Source
- yaak-ai
- Collection Method
- Data likely collected from instrumented vehicles during real-world driving.
- Time Range
- null
- Freshness
- Last updated 2025-09-29 19:30:18; freshness should be verified.
- Geography
- 30 cities in Germany