A new model helps robots understand space and language
The SLA model uses continual pre-training to improve how robots handle tasks in three-dimensional space.
Researchers developed SLA, a Spatial-Language-Action Model. This model helps robots with semantic understanding and precise control in three-dimensional space.
SLA scales implicit spatial alignment through continual pre-training. It is initialized from a pretrained pi05 policy.
The model optimizes action prediction and aligns visual embeddings with spatial representations from a frozen 3D perception model. Feature-dependent gates adaptively aggregate visual embeddings from multiple vision-language-action (VLA) layers.
A teacher uses partially available depth and camera calibration alongside RGB observations. Probabilistic state dropout encourages reliance on visual cues and helps mitigate overfitting to demonstration trajectories.
The training uses curated and unified heterogeneous robot datasets. These support spatial learning across diverse embodiments, tasks, and scenes.
Key facts
- SLA scales implicit spatial alignment to continual pre-training.
- The model uses feature-dependent gates to aggregate visual embeddings from multiple VLA layers.
- SLA preserves the original inference architecture by using auxiliary components only during training.