NVIDIA Research Unveils SpatialClaw, Training-Free Spatial Agent That Beats Prior Agent by 11.2 Points Across 20 Benchmarks

NVIDIA Research introduced SpatialClaw, a training-free agent for complex visual and spatial reasoning tasks that uses code rather than a fixed set of tools as its action interface. The system writes Python inside a persistent kernel and, NVIDIA said, beat a recent prior agent by 11.2 points across 20 benchmarks without benchmark-specific or model-specific tuning.

NVIDIA said the design lets SpatialClaw compose perception modules, inspect intermediate results and revise its strategy across multiple steps, while reusing perception outputs as ordinary variables with libraries such as NumPy and SciPy. The agent also held up consistently across six different model backbones, according to the company.