LeWAM: End-to-End World and Action Modeling with JEPAs
LeWAM claims a single JEPA model can both act directly and plan faster through learned latent dynamics.
The paper introduces joint training for next latent state prediction and action flow matching, with an isotropic Gaussian constraint on the latent state. In visual planning benchmarks, it reports success gains from 28.6% to 89.7% on contact-rich tasks and from 10.9% to 46.2% on long-horizon tasks. It also reports 32.3x faster planning, plus competitive behavior cloning results in simulation and on physical robots. HF Daily Papers' note
The paper introduces joint training for next latent state prediction and action flow matching, with an isotropic Gaussian constraint on the latent state. In visual planning benchmarks, it reports success gains from 28.6% to 89.7% on contact-rich tasks and from 10.9% to 46.2% on long-horizon tasks. It also reports 32.3x faster planning, plus competitive behavior cloning results in simulation and on physical robots. HF Daily Papers' note
score 6