AI Technology: Deep Learning; Vision (image recognition, image generation, text recognition OCR, etc.).
What if each mile driven by a self-driving car enhanced its ability? That vision may be realized through world models, which enable vehicles to learn to predict their environment without requiring human supervision.
Due to their rigid, predefined object classes and costly data annotations, older systems often fail under unpredictable conditions.
World models offer a robust alternative by utilizing self-supervised learning and vast amounts of unlabeled video data.
The project aims to develop openly accessible world models to foster transparency and collaboration in the field, addressing key challenges such as data curation and model scalability.
The initiative seeks to establish a robust, open-access benchmark for driving world models and to democratize access to them.
Significant hurdles have already been overcome, including obtention of data, validation, and benchmarking of our codes. Access to sufficient compute remains a major barrier to overcome, as world models, like most foundation models, remain very expensive to train.
Florent Bartoccioni, Valeo - France