In the previous year, Hugging Face—the AI development platform—unveiled LeRobot, a suite of open AI models, datasets, and resources aimed at facilitating the creation of practical robotics systems. This Tuesday, Hugging Face partnered with AI startup Yaak to enhance LeRobot with a new training dataset designed for robots and vehicles capable of autonomous navigation in various environments, including urban streets.
The newly launched dataset, named Learning to Drive (L2D), surpasses one petabyte in volume. It comprises data sourced from sensors installed in vehicles within German driving schools. L2D encompasses information from cameras, GPS, and vehicle dynamics, documenting the experiences of instructors and students as they maneuver through streets featuring construction sites, intersections, highways, and more.
Numerous open self-driving training datasets exist from various companies, such as Alphabet’s Waymo and Comma AI. However, many of these datasets emphasize planning tasks like object detection and tracking that demand high-quality annotations, as noted by the creators of L2D, which can make them challenging to scale.
Unlike its counterparts, L2D has been tailored to facilitate the development of “end-to-end” learning systems, according to its developers, allowing for actions to be predicted (such as the moment a pedestrian may cross the road) based directly on sensor inputs (like camera video).
“The AI community now has the opportunity to create end-to-end self-driving models,” stated Yaak co-founder Harsimrat Sandhawalia, along with Remi Cadene from Hugging Face’s AI for robotics team, in a blog entry. “L2D is positioned to be the largest open-source self-driving dataset, providing the AI community with unique and varied ‘episodes’ for developing end-to-end spatial intelligence.”
This summer, Hugging Face and Yaak are set to undertake real-world “closed-loop” evaluations of models trained with L2D and LeRobot, using a vehicle supervised by a safety driver. The companies are inviting the AI community to propose models and tasks for evaluation, such as navigating roundabouts and parking scenarios.
Compiled by Techarena.au.
Fanpage: TechArena.au
Watch more about AI – Artificial Intelligence
