LeWAM: End-to-End World and Action Modeling with JEPAs

LeWAM is an end-to-end JEPA framework for control that improves success rates and planning efficiency in both simulated and physical environments. It achieves success rates of 89.7% on contact-rich tasks and 46.2% on long-horizon tasks, with 32.3 times faster planning. The model supports direct policy execution and efficient planning through learned latent dynamics, and is competitive with established methods in behavior cloning tasks.