SLA: Scaling Spatial-Language-Action Model via Continual Pre-training

SLA. Scaling Spatial-Language-Action Model via Continual Pre-training. By Fuhao Li and nine others, from the Hong Kong University of Science and Technology, Guangzhou, OLA Dimensions, and two other institutions. Published on the fifth of November.

A robot that manipulates objects reliably has to understand what it is asked to do and act precisely in three-dimensional space. Those are two different demands, meaning and geometry, and it is the second that concerns the authors. Vision-language-action models, or VLA models, mostly inherit their visual representations from two-dimensional vision-language pre-training. Spatial supervision, when it appears, tends to arrive late, during adaptation to a specific task, rather than as part of the foundation. The authors argue that if spatial understanding is to be a capability the foundation model carries and transfers, it has to be learned from broader robot experience, earlier.

SLA, a Spatial-Language-Action Model, moves spatial alignment into continual pre-training and keeps it in place through downstream post-training. It starts from a pretrained pi05 policy. From there it optimizes two things jointly: predicting actions, and aligning its intermediate visual embeddings with spatial representations produced by a frozen 3D perception model. The alignment is implicit. What gets aligned is the policy's internal representation, matched against the frozen model's spatial features.

Scaling this up meant gathering heterogeneous robot datasets, curated and unified so spatial learning spans many embodiments, tasks and scenes. Data that varied brings its own complications, and three components address them. Feature-dependent gates adaptively combine visual embeddings from several VLA layers before spatial supervision is applied. The teacher makes use of depth and camera calibration where they are available, alongside RGB observations, so uneven geometric annotations still contribute. And probabilistic state dropout pushes the policy to rely on what it sees, which mitigates overfitting to demonstration trajectories.

The abstract reports no benchmark numbers or comparisons. What it does state is a structural property: every auxiliary component is used only during training. Pretrained VLA weights can be reused directly, and the original inference architecture is preserved. Listeners looking for how much the spatial alignment helps on downstream tasks will need the paper itself; the abstract describes the method and stops there.

Here, spatial understanding is something learned at scale and then left behind in the weights. Gates, teacher and dropout all disappear at inference, so the deployed model keeps the architecture it started from. Whatever the spatial training contributed, it contributes through parameters alone.