SLA: Scaling spatial-language-action models via continual pre-training

The study proposes SLA, a spatial-language-action model initialized from a pretrained pi05 policy, which jointly optimizes action prediction and the alignment of intermediate visual embeddings with the spatial representations of a frozen 3D perception model during continual pre-training. The team unified heterogeneous robot datasets spanning embodiments, tasks, and scenes, and introduced feature-dependent gates that adaptively aggregate visual embeddings from multiple layers; the teacher model can use partially available depth and camera calibration information. All auxiliary components are used only during training, and the inference architecture remains unchanged.