Hugging Face Enhances LeRobot Platform with Advanced Training Data for Autonomous Vehicles
Last year, Hugging Face made significant strides in the field of artificial intelligence by launching LeRobot, a comprehensive collection of open AI models, datasets, and tools aimed at facilitating the development of real-world robotics systems. Recently, Hugging Face announced a collaboration with AI startup Yaak to enhance LeRobot with an innovative training set designed specifically for robots and vehicles that can autonomously navigate various environments, such as city streets.
Introducing Learning to Drive (L2D)
The newly introduced dataset, known as Learning to Drive (L2D), is a massive resource exceeding one petabyte in size. It comprises valuable data collected from sensors mounted on vehicles in German driving schools. This dataset captures a variety of information, including:
- Camera footage
- GPS data
- Vehicle dynamics
This extensive data collection includes scenarios involving construction zones, intersections, highways, and other challenging driving conditions, making it an essential tool for AI development.
Comparison with Existing Datasets
While there are several open self-driving datasets available from companies such as Alphabet’s Waymo and Comma AI, many of these focus primarily on high-quality annotations for planning tasks like object detection and tracking. According to the creators of L2D, this focus can complicate scalability.
End-to-End Learning with L2D
In contrast, L2D is specifically designed to foster the development of end-to-end learning. This approach allows for predictions of actions, such as when a pedestrian may cross the street, based directly on sensor inputs like camera footage. Yaak co-founder Harsimrat Sandhawalia and Remi Cadene from the AI for robotics team at Hugging Face expressed in their blog post:
“The AI community can now build end-to-end self-driving models. L2D aims to be the largest open-source self-driving dataset that empowers the AI community with unique and diverse ‘episodes’ for training end-to-end spatial intelligence.”
Future Testing Plans
Hugging Face and Yaak are gearing up for real-world “closed-loop” testing of models trained using L2D and LeRobot this summer. This testing will take place on a vehicle equipped with a safety driver. The companies are inviting the AI community to submit their models and propose tasks for evaluation, such as:
- Navigating roundabouts
- Parking in tight spaces
To stay updated on this collaboration and future developments, check out more about Hugging Face Robotics and join the conversation in the AI community.
