VLAs build on pretrained vision-language backbones and an action head trained on robot demonstration data; they are the common architecture for generalist manipulation policies discussed in robotics papers. Many VLAs are trained mostly from 2D images, so papers such as SLA (arXiv 2611.08373, November 2026) study how to add spatial supervision without changing the inference-time stack. The term is model-family language, not a single vendor product.