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

**Authors:** Fuhao Li, Wenxuan Song, Qizhou Huang, Ziyang Zhou, Jingbo Wang, Junwu E, Pengwei Wang, Yan Wang, Shunbo Zhou, Haoang Li  
**Lab:** HKUSTGZ · **Published:** 5 Nov 2026 · **Upvotes:** 92 · **Level:** Intermediate  
**Links:** [arXiv](https://arxiv.org/abs/2611.08373) · [PDF](https://arxiv.org/pdf/2611.08373) · [This page](https://omen-bit.github.io/ai-paper-daily/paper/2611.08373.html)

## Curator's notes

**TL;DR:** SLA continually pre‑trains a vision‑language‑action model to align visual embeddings with 3D spatial representations, improving transferable spatial understanding for robots.

- **What's new:** Introduces continual pre‑training with implicit spatial alignment and feature‑dependent gates to fuse embeddings across heterogeneous robot datasets, using only training‑time auxiliary components.
- **Why it matters:** Better spatial grounding lets robots generalize manipulation across tasks without redesigning perception.

## Abstract

Reliable robotic manipulation requires both semantic understanding and precise control in three-dimensional space. However, vision-language-action (VLA) foundations largely inherit visual representations from 2D vision-language pre-training, while spatial supervision is often introduced only during task-specific adaptation. Developing spatial understanding as a transferable foundation-model capability therefore calls for learning from broader robot experience. We present SLA, a Spatial-Language-Action Model that scales implicit spatial alignment to continual pre-training and maintains it during downstream post-training. Initialized from a pretrained pi05 policy, SLA jointly optimizes action prediction and alignment between intermediate visual embeddings and spatial representations from a frozen 3D perception model. We curate and unify heterogeneous robot datasets to support spatial learning across diverse embodiments, tasks, and scenes. To accommodate this diversity, we introduce feature-dependent gates that adaptively aggregate visual embeddings from multiple VLA layers for spatial supervision. The teacher further exploits partially available depth and camera calibration alongside RGB observations, enabling effective use of heterogeneous geometric annotations. Probabilistic state dropout encourages reliance on visual cues and mitigates overfitting to demonstration trajectories. All auxiliary components are used only during training, allowing direct reuse of pretrained VLA weights and preserving the original inference architecture.

## What you could build

1. Create a robot pick‑and‑place system that uses SLA to plan grasps in unseen scenes.
2. Develop a simulation‑to‑real transfer pipeline leveraging SLA’s spatial embeddings.
3. Build an interactive demo where a robot follows natural language commands to navigate cluttered environments.

## Quiz yourself

What auxiliary components does SLA use only during training to preserve the original inference architecture?

**Answer:** Feature‑dependent gates, probabilistic state dropout, and a teacher that exploits depth and camera calibration are used only during training.

---

Notes written by AI from the paper, numbers copied from the abstract. Check the paper before you rely on them. From AI Papers Daily.