Research
ALPS, a visual state encoder unifying physical state, language and reconstruction for world models
ALPS is a ViT-B/16 pretrained on still images and text with language, reconstruction and spatial objectives. It reports 85.5% mean LIBERO policy success with its native text tower and matches V-JEPA 2-L on aggregate state-probe error despite being 3.5x smaller.
Every outlet that ran it 1 article
Hugging Face PapersResearchLead