Hugging Face, a prominent platform in AI development has made strides in expanding its LeRobot initiative, which encompasses a collection of open AI models and tools aimed at advancing robotics systems. In a recent partnership with AI startup Yaak, Hugging Face announced the introduction of a new training dataset specifically for autonomous robotics and vehicles.
This dataset, named Learning to Drive (L2D), exceeds a petabyte in size and utilizes data sourced from vehicles fitted with sensors in German driving schools. This dataset captures essential information such as camera images, GPS coordinates, and vehicle dynamics as driving instructors and students navigate various real-world settings, including complicated scenarios like construction zones and busy intersections.
Unlike existing self-driving training datasets from competitors such as Alphabet’s Waymo and Comma AI, which largely focus on high-precision tasks like object detection and tracking, L2D is aimed at facilitating “end-to-end” learning.
This approach enables models to predict actions based directly on sensor input, such as determining when a pedestrian might cross the street solely from camera footage.
According to Yaak co-founder Harsimrat Sandhawalia and Remi Cadene from Hugging Face’s AI for Robotics team, this innovative dataset is intended to be the most prominent open-source resource for developing end-to-end self-driving models, offering unique and varied scenarios or “episodes,” that foster the training of spatial intelligence.
Looking ahead, Hugging Face and Yaak have plans to perform real-world closed-loop testing of the models trained on L2D and LeRobot later this summer, utilizing a vehicle equipped with a safety driver.
They invite the AI community to contribute models and suggest tasks for assessment, such as navigating through roundabouts and parking efficiently. This initiative marks a significant leap forward in the effort to enhance autonomous vehicle capabilities through collaborative data-sharing and innovation in AI development.
