HyVLA-0.5

HyVLA-0.5 (Hy-Embodied-0.5-VLA) is a Vision-Language-Action model that constitutes the real-world robot policy component of the HY-Embodied project. It builds upon the HY-Embodied-0.5-MoT backbone (4B parameters), adding a flow-matching action expert, a compact memory encoder with temporal-spatial attention, and a delta-chunk action representation. It is pre-trained on Hy-UMI-10K, a 10,000-hour egocentric UMI dataset, and refined with preference optimization. Evaluated on RoboDojo benchmark among 30 policies.