VLA-JEPA: Training Robots to Focus on Latent Action Space

jiqizhixin · x · 2026-07-04

Research teams from USTC, Tsinghua University, and SJTU have proposed VLA-JEPA. By predicting future states in the latent space rather than at the pixel level, it avoids common issues like appearance bias and camera movement. It demonstrates better generalization and robustness than existing VLA methods on LIBERO, SimplerEnv, and real-world manipulation tasks. The arXiv paper, code, and project page are now publicly available.

Original post →

More from Embodied

Embodied channel →