On Wednesday, Hugging Face revealed an expansion of its LeRobot platform with the introduction of a substantial dataset designed for automotive automation. This dataset, known as Learning to Drive (L2D), was developed in partnership with the AI startup Yaak. Over a three-year period, data was gathered from a network of sensors installed on 60 electric vehicles (EVs). The goal of this open-source dataset is to empower developers and the robotics sector to create innovative spatial intelligence solutions within the automotive field.
Hugging Face Introduces L2D Dataset to LeRobot
In a blog post, the company highlighted the new AI dataset, referring to it as “the world’s largest multimodal dataset aimed at building an open-sourced spatial intelligence for the automotive domain.” The dataset spans over 1 petabyte (PB) in size and was compiled from sensor data across 60 EVs operated by driving schools in 30 different cities in Germany. To maintain consistency, identical sensor setups were employed throughout the data collection process.
Launched last year, the LeRobot platform serves as a repository for open-source AI models, datasets, and essential tools to assist developers in creating AI-driven robotic systems.
The Learning to Drive dataset
Photo Credit: Hugging Face
The dataset’s policies are segmented into two categories: expert policies and student policies. Expert policies consist of data collected from professional driving instructors, while student policies are derived from learner drivers. According to Hugging Face, expert policies demonstrate flawless driving performance, while student policies contain discernible sub-optimal driving behaviors. Both categories include natural language instructions for various driving tasks.
Each policy group encompasses all the driving scenarios necessary for obtaining a driving license in the European Union (EU). These scenarios include key tasks such as overtaking, navigating roundabouts, and driving on tracks.
In discussing the sensor technology utilized for the L2D dataset, Hugging Face noted that each of the 60 Kia Niro EVs was outfitted with six RGB cameras, which provided a 360-degree view of the vehicle’s environment. Additional equipment included on-board GPS for location tracking and mapping, along with an inertial measurement unit (IMU) to monitor vehicle dynamics. All collected data was accurately timestamped.
Importantly, the dataset intends to facilitate the development of comprehensive self-driving AI models suitable for use in fully autonomous vehicle systems.
Hugging Face indicated that the L2D dataset will be released in phases, with each subsequent release being an enhancement of its predecessor to ensure user accessibility. Additionally, the platform welcomes the community to contribute models for closed-loop testing of the dataset, which will include a safety driver, scheduled to commence in summer 2025.
