Friday, 6 November 2026 · 7 min read
SLA improves robotic manipulation by continual pre-training with 3D spatial alignment
SLA enhances vision-language-action models by aligning visual embeddings with 3D spatial features during continual pre-training and post-training, achieving up to 87.4% success on LIBERO-Plus and 24.0 percentage points gain over baseline in real-world tasks.
Based on the paper SLA: Scaling Spatial-Language-Action Model via Continual Pre-training by Fuhao Li, Wenxuan Song, Qizhou Huang, Ziyang Zhou and 6 others
Challenges in integrating 3D spatial understanding into vision-language-action models
Robotic manipulation requires combining semantic understanding from vision and language with precise 3D spatial control. Vision-language-action (VLA) models typically build on 2D vision-language pretraining, lacking explicit 3D spatial knowledge during pretraining, which limits their ability to reason about spatial relationships crucial for manipulation.
Existing approaches add 3D information during task-specific post-training by incorporating depth or point cloud data, but these methods face challenges: 3D sensor data can be noisy or incomplete, sensors vary widely across robots, and many large robot datasets lack depth annotations, limiting scalability.
Spatial Forcing introduced implicit 3D alignment during post-training but left spatial understanding as a task-specific add-on. Developing a scalable 3D-aware foundation model requires integrating spatial learning during continual pre-training on diverse robot data, preserving pretrained knowledge while enhancing spatial reasoning.
Sign in to read the whole paper
You have read the opening. The rest is free with a Bezen account, and no card is needed.
In the rest of this paper
- How the method works, step by step, with a flow chart
- Three key formulas from the paper, explained symbol by symbol
- Three figures from the paper, and how to read them
- How it was tested, what was found, and where it falls short