Researchers add 3D spatial supervision to robot VLA pre-training

A paper on arXiv presents SLA, a Spatial-Language-Action model that builds 3D spatial understanding into continual pre-training of a robot policy rather than leaving it to task-specific adaptation. Starting from the pretrained pi05 policy, SLA optimizes action prediction together with alignment between its intermediate visual embeddings and spatial representations from a frozen 3D perception model. The authors unify heterogeneous robot datasets covering different embodiments, tasks and scenes, and add feature-dependent gates that combine visual embeddings from several VLA layers for supervision. The teacher makes use of whatever depth and camera calibration exists alongside RGB, and probabilistic state dropout pushes the model toward visual cues. The auxiliary components are training-only, so inference keeps the original VLA architecture.