LeWAM: Lightweight World Action Model Hits 92.28% Success on RoboTwin 2.0

udmrzn · x · 2026-09-25

Researchers introduced LeWAM (Latent evolving World Action Model), a lightweight world action model that unifies action generation and future visual prediction in I-JEPA's latent space without video diffusion. With AdaFuse and DemoDPO, it achieves a 92.28% average success rate on the RoboTwin 2.0 benchmark, showing latent-space prediction is a competitive alternative to video-diffusion-based world models for robot manipulation.

Original post →

More from Embodied

Embodied channel →