arXiv 2611.08373
SLA: Scaling Spatial-Language-Action Model via Continual Pre-training
Robotics48 upvotes
Authors: Fuhao Li, Wenxuan Song, Qizhou Huang, Ziyang Zhou, Jingbo Wang, Junwu E, Pengwei Wang, Yan Wang +2
Overview
The authors present SLA, a Spatial-Language-Action model initialized from a pretrained pi05 policy. During continual pre-training, it jointly optimizes action prediction and alignment between intermediate VLA visual embeddings and features from a frozen 3D perception model, using feature-dependent gates and probabilistic state dropout. All auxiliary components are used only during training, so the original inference architecture is kept.
Introduction
Robotic manipulation needs both semantic understanding and precise 3D control, but vision-language-action models mostly inherit 2D visual representations and add spatial supervision only during task-specific adaptation. Spatial understanding therefore needs to be built into the foundation model itself.
Results
The abstract does not report quantitative results. It states that SLA's spatial alignment is maintained during downstream post-training and that pretrained VLA weights can be reused directly.
Conclusion
Building spatial alignment into VLA pre-training, without changing inference architecture, offers a practical way to give robot foundation models transferable 3D understanding.