← Back to live feed1 story
Meta trained an 8 billion parameter JEPA model that demonstrates robot world models follow scaling laws similar to large language models, and AMI released H-JEPA, a hierarchical system using three stacked JEPAs that improved success on the Visual Antmaze benchmark from 18% to 73%.
Sign in to suggest edits
Key sources
- SOURCE@arankomatsuzaki“Raises success from 18% to 73% on Visual AntMaze, while using less planning compute.”x.com
- SOURCE@basileterv987“Hierarchical world models, learned end-to-end from pixels!”x.com
- SUPPORT@randall_balestr“the first end-to-end learned hierarchical world model for long-horizon visual planning!”x.com
- SOURCEhuggingnewshuggingnews.com