NVIDIA Research Bets on Code, Not Tool Calls, to Fix AI Spatial Reasoning
DevOps.com, Monday, June 22nd, 2026
NVIDIA's SpatialClaw lets AI agents write code to reason about 3D space, boosting accuracy.
NVIDIA Research released SpatialClaw, an open-source, training-free framework that rethinks how AI agents determine where things are in physical space.
It targets a weakness in vision-language models, which describe scenes well but struggle with geometry like distance, orientation, and motion across frames.
Instead of calling a fixed menu of tools, the agent writes Python in a persistent kernel so it can compose perception modules, inspect results, and revise strategy.
Across 20 spatial reasoning benchmarks it averaged 59.9% accuracy, an 11.2-point improvement over the previous best spatial agent.