AMI Launches First Hierarchical World Model for Long Horizon Visual Planning

AMI launched H-JEPA, which it describes as the first end-to-end learned hierarchical world model for long horizon visual planning. The model improved success on the Visual Antmaze benchmark from 18% to 73% while reducing planning compute, learning world models at various abstraction levels and planning from coarse goals to finer actions. The architecture is built atop SIGReg/LeWM and uses a stable training recipe that allows semantic abstractions to emerge naturally across its levels.