SLA adds 3D spatial alignment to robot policy pre-training without changing inference
The SLA paper continues pre-training a pi05 policy with alignment to a frozen 3D perception model, then drops the extra parts so the inference architecture stays unchanged.
The problem
Reliable robotic manipulation needs both semantic understanding and precise control in three-dimensional space. The authors argue that vision-language-action (VLA) foundation models mostly inherit their visual representations from 2D vision-language pre-training, while spatial supervision is often added only during task-specific adaptation. Their position is that spatial understanding should be a transferable capability of the foundation model, which calls for learning it from broader robot experience.
The paper, SLA: Scaling Spatial-Language-Action Model via Continual Pre-training (arXiv 2611.08373), lists 10 authors and the organization HKUSTGZ. It was published on arXiv on 2026-11-05 and listed on Hugging Face daily papers on 2026-11-04.
What SLA does
SLA, a Spatial-Language-Action Model, is initialized from a pretrained pi05 policy. During continual pre-training it jointly optimizes two things: action prediction, and alignment between intermediate visual embeddings and spatial representations from a frozen 3D perception model that serves as the teacher. The authors describe this as scaling implicit spatial alignment to continual pre-training, and they keep the alignment in place during downstream post-training as well.
To give the model enough spatial signal, the authors curate and unify heterogeneous robot datasets that span diverse embodiments, tasks and scenes. Three pieces deal with that diversity:
- Feature-dependent gates adaptively aggregate visual embeddings from multiple VLA layers for spatial supervision.
- The teacher uses depth and camera calibration where they are partially available, alongside RGB observations, so datasets with different geometric annotations can all contribute.
- Probabilistic state dropout encourages the policy to rely on visual cues and mitigates overfitting to demonstration trajectories.
All of these auxiliary components are used only during training. According to the abstract, that allows direct reuse of pretrained VLA weights and preserves the original inference architecture. The spatial teacher, the gates and the dropout scheme are not part of the model that runs at inference time.
What the abstract leaves open
The abstract reports no results. It names no benchmarks, simulators or real-robot tasks, gives no success rates, and makes no comparison against the pi05 starting point or other VLA models. It does not say which robot datasets were unified or how large the combined collection is, and it does not name the frozen 3D perception model used as the teacher. It also leaves open how much each piece contributes on its own: the gates, the use of depth and calibration, and state dropout. No code repository is listed with the paper. Readers who want those details will need the full paper on arXiv.
Source: SLA: Scaling Spatial-Language-Action Model via Continual Pre-training (arXiv 2611.08373)