DEV Community

Cover image for SLA Adds 3D Spatial Alignment to VLA Pre-Training, Reaching 83.8% Real-World Success
aimodels-fyi
aimodels-fyi

Posted on Originally published at aimodels.fyi

SLA Adds 3D Spatial Alignment to VLA Pre-Training, Reaching 83.8% Real-World Success

This is a Plain English Papers summary of a research paper called SLA Adds 3D Spatial Alignment to VLA Pre-Training, Reaching 83.8% Real-World Success. If you like these kinds of analyses, you can find more AI and machine-learning research on AIModels.fyi or follow us on Twitter.

SLA moves spatial learning into VLA pre-training

SLA starts from the pretrained pi0.5 policy and adds spatial alignment during continual pre-training, then keeps that objective during downstream post-training. Its central technical move aligns intermediate visual embeddings with pixel-level 3D representations from frozen OmniVGGT, while the policy also learns to predict robot actions. The auxiliary teacher and alignment modules are discarded at deployment.

The motivation is a gap in common VLA training: visual features inherit much of their structure from 2D image-text pre-training, while geometric supervision often arrives only during task adaptation. SLA instead uses broad robot experience to make spatial information a transferable capability. The authors report 87.4% on LIBERO-Plus and an average 83.8% success rate across eight real-world deployment conditions. SLA’s focus on spatial representations connects to work such as SpatialVLA, which also studies spatial representations for VLA manipulation.

How frozen 3D supervision shapes the policy...


Continue reading the full paper summary on AIModels.fyi →

Top comments (0)