Paper review · November 7, 2026
SLA: A Spatial-Language-Action Model Extended Through Continual Pretraining
SLA: Scaling Spatial-Language-Action Model via Continual Pre-training
SLA continues pretraining the π0.5 VLA on mixed robot data while aligning its visual representations with features from a frozen 3D perception model. This improves spatial understanding and manipulation performance without changing the inference architecture.
Problem
VLA foundation models inherit their visual representations from 2D image-text pretraining. Spatial information is usually added only during task-specific fine-tuning. Adding depth sensor input often runs into noisy or missing data. It also changes the model architecture, which can damage knowledge learned in pretraining.
Method
Starting from a pretrained π0.5 policy, the method adds an alignment loss to the action prediction loss. This loss raises the cosine similarity between the VLA's intermediate visual tokens and features from a frozen 3D perception model (OmniVGGT). The model goes through continual pretraining on a training corpus that unifies the state and action formats of several public robot datasets, and the alignment is kept during later fine-tuning. A feature-dependent gate adaptively combines visual embeddings from several layers. Depth and camera calibration information, which only part of the data has, is also given to the teacher model. Stochastic state dropout reduces overfitting to demonstration trajectories. The auxiliary modules are used only during training, and the inference architecture stays the same.
Results
- On LIBERO-Plus, the authors report an overall success rate of 87.4% under observation, language and scene perturbations, higher than the compared baselines.
- On RoboTwin 2.0, trained only on clean demonstrations, it reached a 72.3% success rate in the Clean2Random setting.
- On RoboCasa-Panda, it reached 68.4% with only 50 demonstrations per task. The authors report that this is higher than comparison policies trained on 300–3,000 demonstrations per task.
- On RoboDojo-Real, a single policy adapted to three robots scored 33.9%, well ahead of π0.5 (12.8%).
- Across eight real-world deployment conditions, it averaged 83.8%, 24.0 percentage points higher than π0.5. Aligning in both training stages beat single-stage alignment at every data ratio.
Limitations
Only one base model, π0.5, was tested. The quality of the spatial signal depends on the frozen 3D teacher (OmniVGGT). The provided text includes only part of the ablations on each component's contribution and part of the detailed numbers for each baseline. The real-world evaluation is limited to eight conditions.
Why it matters
3D spatial understanding can be improved without depth sensors or extra modules, while keeping the existing VLA weights and inference architecture. This makes it easier to upgrade robot policies that are already deployed. It also performs well with few demonstrations, which can reduce the burden of collecting field data.
Key terms
- VLA(Vision-Language-Action)
- A model that takes images and language instructions as input and directly outputs robot actions.
- Continual pretraining
- A stage in which a model that is already pretrained is trained further on a large amount of additional data to build specific abilities while keeping what it already knows.
- Representation alignment
- A method that adds an auxiliary loss so that a model's internal features become similar to the features of an external teacher model. The teacher is not needed at inference.
- State dropout
- A technique that randomly removes the robot state input during training so the model relies more on visual cues.
Original
Authors: Fuhao Li, Wenxuan Song, Qizhou Huang, Ziyang Zhou, Jingbo Wang, Junwu E, Pengwei Wang, Yan Wang, et al. · Posted 2026-11-05 · License arXiv 비독점 배포 라이선스 · Based on the full text
This review is an AI-written summary of the paper. Please treat the original paper as the reference.