SLA
Research project Spatial-language-action model

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

Fuhao Li1,*Wenxuan Song1,*,†Qizhou Huang1Ziyang Zhou1Jingbo Wang1Junwu E1Pengwei Wang3Yan Wang4Shunbo Zhou2Haoang Li1,2

1The Hong Kong University of Science and Technology (Guangzhou) · 2OLA Dimensions
3Beijing Academy of Artificial Intelligence · 4Tsinghua University
*Equal Contribution · †Project Lead

↗ arXiv2611.08373 ↗ PDFRead the paper ↗ GitHubCode repository ↗ Hugging FaceSLA collection
Contents

01 / Introduction

Spatial understanding at foundation-model scale

Reliable robotic manipulation requires not just semantic understanding but even more precise control. SLA brings spatial supervision into continual pre-training, then reinforces it during downstream post-training.

SLA overview showing heterogeneous robot data, intermediate visual feature alignment with a frozen 3D model, and evaluation results
Figure 1. A unified path from curated robot data to spatially aligned visual representations and robust manipulation.

Starting from a pretrained π0.5 policy, the model aligns intermediate visual embeddings with representations from a frozen 3D perception model while learning robot actions. The external teacher and alignment modules are used during training; inference retains the original policy architecture.

02 / Training data

Learning from heterogeneous robot experience

The continual pre-training corpus combines robot demonstrations across embodiments, tasks, and scenes. A unified processing pipeline reconciles action conventions and available geometric annotations before training. The policy then adapts to each downstream benchmark through post-training.

Table 1. Overview of the source datasets for continual pre-training. Statistics refer to the source datasets before our data selection and processing; task counts are summed across datasets.
Data sourceEpisodesTasksHours
InternData-A1630K707430
AgiBotWorld-Beta168K2172970
RoboCasa365-Pretrain630K3002090
RoboCOIN183K8081300
DROID95K86510
MolmoAct7.9K2330
Total1,713.9K1,50414,330
Continual pre-trainingSpatial alignment + action learning
→
Task post-trainingBenchmark and real-world adaptation
→
DeploymentOriginal VLA inference architecture

03 / Evaluation

Results across simulation and real robots

The paper evaluates robustness, scene generalization, downstream data efficiency, and deployment across robot embodiments. The figures below report success rates from the current manuscript.

LIBERO-Plus87.4%

Weighted success rate across seven perturbation dimensions

RoboTwin 2.072.3%

Clean2Random success rate across 50 tasks

RoboCasa-Panda68.4%

Mean success rate across 24 kitchen tasks

RoboDojo-Real33.9%

Overall success rate across three embodiments

Data scaling curves comparing spatial alignment in continual pre-training and post-training
Figure 2. Using spatial alignment during both training stages gives the strongest RoboTwin Clean2Random results at each tested data scale.

04 / Deployment

Generalization under real-world changes

Eight deployment conditions test manipulation under human perturbations, moving targets, table-height changes, and instruction-guided handling of an unseen object. Across these conditions, SLA improves average success over π0.5 by 24.0 percentage points.

Base: Robot grasps five vegetables between both hands

Position Generalization: Human throw fruit out, and robot picks it back up

Dynamic: Human intentionally move fruit while robot pick it

Level 1: Grasp water bottle

Level 2: Grasp watch from stationary table

Level 3: Grasp watch from dynamic table

Instruction-following: Grasp black tape

05 / Reference

Citation

@article{li2026sla,
  title={SLA: Scaling Spatial-Language-Action Model via Continual Pre-training},
  author={Li, Fuhao and Song, Wenxuan and Huang, Qizhou and Zhou, Ziyang and Wang, Jingbo and E, Junwu and Wang, Pengwei and Wang, Yan and Zhou, Shunbo and Li, Haoang},
  journal={arXiv preprint arXiv:2611.08373},
  year={2026}
}