La-JEPA preprint reports faster robot control with a smaller policy model

Researchers report that La-JEPA, a compact robot control model, runs substantially faster than Fast-WAM at comparable performance. The preprint also reports average success comparable to pi0.5 on four real-robot tasks. La-JEPA trains a policy transformer from scratch using features from a pretrained V-JEPA 2.1 encoder. During training, it predicts future features alongside actions; during operation, it generates only actions. It avoids a pretrained generative policy backbone. The researchers also corrected instructions corrupted by the official LIBERO-Plus evaluator and re-evaluated several baselines. Their state-of-the-art results on that benchmark and RoboTwin 2.0 are limited to policies without embodied pretraining.