Breaking

NVIDIA Releases SpatialClaw, a Code-Driven Spatial Reasoning Agent

June 16, 2026 at 16:01 EDT

AI Agents
Foundation Models
Open Source

NVIDIA Research has released SpatialClaw, a training-free spatial reasoning agent that writes Python code to solve complex visual tasks instead of calling a fixed set of predefined tools. The accompanying paper was submitted to arXiv on June 11, 2026, and the code is available as open source on GitHub and a project page.

NVIDIA Research · SpatialClaw

When the agent's tool is code itself: a training-free leap in 3D spatial reasoning

Instead of calling a fixed set of tools, SpatialClaw writes and runs Python inside a stateful kernel — inspecting masks and depth maps, then revising strategy. No training, no fine-tuning, just code as the action interface.

59.9%

Avg. accuracy across 20 benchmarks (Gemma 4-31B)

19/20

Benchmarks improved over the no-tool baseline

+11.2pp

Over prior spatial agent SpaceTools-Toolshed

The action interface matters

Average accuracy gain over the no-tool baseline (53.4%), across 20 benchmarks.

No-tool
53.4%

+1.8

Single-pass code
55.2%

+3.3

Structured tool-call
56.7%

+6.5

SpatialClaw
59.9%

Code as the action interface — an iterative loop

1 · WRITE

VLM writes a Python cell into a persistent kernel

→

2 · RUN

SAM 3 masks & Depth-Anything-3 become variables; NumPy/SciPy compute

→

3 · INSPECT & REVISE

Read intermediate results, adapt strategy, repeat

Largest per-benchmark gains

Where iterative spatial reasoning pays off most (points improved).

+17.6

DSI-Bench (4D)

+15.3

MindCube (multi-view)

+13.4

MMSI

Consistent +3.1 to +7.7 point gains held across six VLM backbones (26B–397B params, Qwen3.5/3.6 & Gemma4).

Why developers like it

Stepping through a stateful kernel — checking masks and depth maps as it goes — and decomposing tasks into detection, coordinates, distance and orientation. Strong promise for robot vision and Physical AI.

The caveat

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.