NVIDIA Introduces SpatialClaw for Training-Free Spatial Reasoning

NVIDIA Research has unveiled SpatialClaw, a training-free framework designed to address existing weaknesses in vision-language models (VLMs) when it comes to spatial reasoning. Instead of retraining models, SpatialClaw treats code as the action interface, allowing agents to dynamically compose tools and inspect results to refine their understanding of object relationships and movement in 3D. It achieves an average accuracy of 59.9% across 20 benchmarks, outperforming previous approaches by improving vision-language understanding.