Study proposes SLA spatial-language-action model

The research team proposes the SLA spatial-language-action model, initialized from the pretrained pi05 policy. During continual pretraining, it simultaneously optimizes action prediction and the alignment between intermediate visual embeddings and the spatial representations of a frozen 3D perception model, and maintains this alignment into the downstream post-training stage. During training, in addition to RGB observations, the teacher model also uses partially available depth and camera calibration information to leverage heterogeneous geometric annotations; the team also unified heterogeneous robot datasets across embodiments, tasks, and scenes, and introduced feature-correlation gating to adaptively aggregate multi-layer visual embeddings. All auxiliary components are used only during the training stage, and the inference architecture remains unchanged, so the pretrained VLA weights can be reused directly.

AI 根据报道生成 · 5 days ago更新