SpatialClaw for Inspection

A step-by-step build of a code-as-action agent for image-only inspection, inspired by NVIDIA’s SpatialClaw: what the paper really does, what to keep and drop, and how to test whether it beats a plain tool-call agent.
series
agentic-ai
computer-vision
inspection
Author

Hasan Goni

Published

October 26, 2026

Modified

October 26, 2026

0.1 Series Overview

NVIDIA’s SpatialClaw lets a vision-language model write Python, one cell at a time, in a notebook that remembers everything. It is built for 3D spatial questions on photos and video. Inspection systems have a different job: one image, a known scale, and a verdict someone is accountable for.

This series is about taking the idea and rebuilding it for that job, in public, with every number measured on data I generate myself. It is deliberately not part of the agentic development series: that one teaches how agents work; this one asks whether a specific new idea earns a place in a production inspection stack.

Two rules for the whole series:

  • Run it, or say so. Every code block that claims a result shows real captured output. Stand-ins are labelled. Anything not run is stated as not run.
  • Licences first. SpatialClaw’s code is non-commercial research only. This series rebuilds the design from the paper with my own code and tracks the licence of every piece.

0.2 Roadmap

Part Post Status
1 What SpatialClaw actually does: paper, code, two real runs, weaknesses Written (not yet published)
2 The adaptation plan: keep, swap, drop, licences, the experiment Written (not yet published)
3 The smallest kernel loop: my own code, a local 27B model, 12 real images Written (not yet published)
4 Eyes and hands for the loop: show(), four tools, three set-ups, each run twice Written (not yet published)
5 A real sandbox, and trying to break it: container, 634-cell replay, 15 attacks Written (not yet published)
6 The A/B/C experiment: detector vs tool-call agent vs code-kernel agent Planned
7 Typed verdicts, the rule layer, audit log, human review Planned
8 Production patterns on real inspection tasks Planned

Links go live as each part is written, in order. Each post will have a narrated animation in the same warm voice as the rest of the blog.

Reuse

Citation

BibTeX citation:
@online{goni2026,
  author = {Goni, Hasan},
  title = {SpatialClaw for {Inspection}},
  date = {2026-10-26},
  langid = {en}
}
For attribution, please cite this work as:
Goni, Hasan. 2026. “SpatialClaw for Inspection.” October 26.
Free email course · 7 days

Industrial Vision Foundations

AI can write the code, but someone still has to know which method fits a real inspection line. One short lesson a day for 7 days on the classical computer vision techniques that still run production systems.

Runnable code in every lesson. No spam, unsubscribe any time.

Questions, corrections or feedback?

Found a mistake, or want a topic covered next? Leave a comment below (sign in with GitHub), or write to hasanme1412@gmail.com. I read everything.