Fuhao Li, Wenxuan Song, Qizhou Huang, Ziyang Zhou, Jingbo Wang, Junwu E, Pengwei Wang, Yan Wang (HKUST-GZ) · Spatial VLA Pretraining

SLA: Scaling Spatial-Language-Action Model via Continual Pre-training

SLA is a spatial-language-action model that builds on a pretrained pi05 robot policy and adds continual pre-training to align intermediate visual embeddings with spatial representations from a frozen 3D perception model. It uses feature-dependent gates over multiple VLA layers, uses partial depth and camera calibration during training, and applies probabilistic state dropout. The auxiliary components are used only during training, so pretrained VLA weights can be reused with the original inference architecture. The page text gives no benchmark numbers or licence details.