single mention · data curation
Multimodal web-data co-training
Multimodal web-data co-training trains a robot model on web data alongside its robot action data.1 The web data contributes broad knowledge of physical objects, letting the model follow language commands involving object categories it has not seen in robot data.2
How this definition was written
Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.
additionally uses HL and WD data
— π0.5\pi_{0.5}: a Vision-Language-Action Model with Open-World Generalizationtraining with web data, which contains very broad knowledge of physical objects, allows the model to understand and follow language commands involving unseen object categories
— π0.5\pi_{0.5}: a Vision-Language-Action Model with Open-World Generalization
Also called Multi-modal Web Data (WD).
Filed under data curation › data mixture & curriculum