@Phoenixyin13: NVIDIA's SpatialClaw is fresh out. This framework directly lets VLM write code step by step in a persistent Python environment, like Jupyter. From calling SAM3 to see things, compute depth, use NumPy and SciPy to process data, view results in real time, if it doesn't work…
Summary
NVIDIA has launched SpatialClaw, a code-based training-free agent framework for complex visual-spatial reasoning tasks, achieving an average of 59.9% on 20 benchmarks, 11.2 points higher than the previous best model.
View Cached Full Text
Cached at: 06/17/26, 08:03 PM
NVIDIA’s SpatialClaw just dropped.
This framework lets VLM directly write code step by step in a persistent Python environment, like a slow-paced Jupyter. It starts by using SAM3 to see things, computes depth, processes data with NumPy and SciPy, views results in real time — if something’s off, just edit and iterate. Incredibly efficient.
As a result, the average score on 20 spatial reasoning benchmarks is 59.9%, a full 11.2 points higher than the previous best, SpaceTools — especially excelling on complex tasks like multi-view, video, and 4D.
“Code is the right action interface for spatial reasoning agents.”
A truly flexible workbench that opens up a new world, letting you push your limits to the max.
NVIDIA AI (@NVIDIAAI): Code is the right action interface for spatial reasoning agents.
New from NVIDIA Research: SpatialClaw, a training-free agent that uses code as its action interface for complex visual tasks.
Instead of calling a fixed set of pre-defined tools, the agent writes Python inside a
Similar Articles
@HuggingPapers: SpatialClaw NVIDIA drops a training-free spatial reasoning agent that uses code as its action interface. A VLM writes P…
NVIDIA introduces SpatialClaw, a training-free spatial reasoning agent that uses a VLM to write Python code in a persistent kernel, compose perception tools, and revise plans, achieving +11.2 points over prior agents on 20 benchmarks.
SpatialClaw: Rethinking Action Interface for Agentic Spatial Reasoning
SpatialClaw is a training-free framework that uses code as an action interface to enable flexible, stateful spatial reasoning in vision-language models, achieving superior performance across diverse 3D/4D spatial reasoning tasks.
SpatialCLI: Learning to Reason With Spatial Tools, Then Without Them
SpatialCLI presents a framework that trains vision-language models to use specialist spatial tools and then internalize those capabilities, boosting Qwen3-VL-8B-Instruct on the MindCube benchmark from 29.3% to 84.6% with tools.
VisualClaw: A Real-Time, Personalized Agent for the Physical World
VisualClaw is a self-evolving multimodal agent that reduces deployment costs through hybrid encoding and skill evolution, while improving video-QA accuracy across multiple benchmarks.
@xiaogaifun: https://x.com/xiaogaifun/status/2073771786202939572
The ByteDance Seed team released the EdgeBench benchmark, which allows AI models to work continuously for 12-72 hours to evaluate their learning ability in long-horizon tasks. They discovered that the relationship between the time spent learning from the environment and performance follows a log-sigmoid curve, revealing a new scaling law.