NVIDIA Research releases SpatialClaw, a training-free spatial reasoning framework
Original title:NVIDIA AI Introduce SpatialClaw: A Training-Free Agent That Treats Code as the Action Interface for Spatial Reasoning
AI summary
NVIDIA Research releases SpatialClaw, a training-free spatial reasoning framework. By using code as an action interface, it lets agents call perception tools (Depth Anything 3, SAM 3) and freely compose their outputs, addressing vision-language models' weaknesses in 3D spatial judgment. Across 20 benchmarks it achieves 59.9% average accuracy, 11.2 percentage points above the recent agent SpaceTools, 6.5 points above a no-tool baseline and 3.2 points above structured tool calling. The framework requires no retraining, and the same prompt and tool set run across all benchmarks and backbones, supporting models from 26B to 397B parameters including Qwen3.5/3.6 and Gemma4.
Key Takeaways
- NVIDIA Research releases SpatialClaw,a training-free spatial reasoning framework.
- By using code as an action interface,it lets agents call perception tools (Depth Anything 3,SAM 3)…
- Across 20 benchmarks it achieves 59.
- 9% average accuracy,11.
- 2 percentage points above the recent agent SpaceTools,6.
Quick Facts
- Published
- Updated
- Source
- MarkTechPost(RSS)
- Category
- Products & Apps
- AI score
- 77
NVIDIA 把空间推理的动作接口从工具调用换成代码,这个思路很巧,20 个基准平均拉升到 59.9%,无训练即插即用,做机器人和视频理解的人该直接跑一下 repo。
Related entities
Source & copyright
This article is based on curated titles and AI summaries from the AIHOT public API, with attribution and links back. Original copyright remains with the original publishers and authors; this site does not mirror full texts.
How to Cite
When citing this article, keep the title, date, and permalink, and credit Orange News as the source.
Orange News: NVIDIA Research releases SpatialClaw, a training-free spatial reasoning framework. June 20, 2026. https://new.nat6.net/en/news/cmqlj0jn600pdsljggy6btum5/