NVIDIA open-sources SpatialClaw: code as action interface lifts spatial agents by 11.2 points
yshan2u · x · 2026-09-19
NVIDIA and KAIST released SpatialClaw, an open-source, training-free framework for spatial reasoning agents built on one idea: code is the right action interface.
- A VLM-backed agent writes Python in a persistent kernel, composes perception outputs, inspects them, and revises — no extra training required
- Evaluated on 20 benchmarks across 6 open-source VLMs, averaging 59.9%, an 11.2-point improvement over prior spatial agents
- Code is available at NVlabs/SpatialClaw; the team argues the action interface matters as much as the backbone model
The thread also touches on academia's role in the era of large models, with the earlier rejected VIGA inverse-graphics work as a precursor.
Related event: NVIDIA Open-Sources Training-Free Spatial Reasoning Agent SpatialClaw(2 posts)→
More from coding & agent
- Econ PhD Job Market Database Hits 900+ Candidates, Open to Codex and Claude Code — aniketapanjwani · 2026-11-02
- Cron works by hand, fails on schedule: a 7-point debugging checklist — petrusenko_max · 2026-11-02
- Local quants vs Claude on 9 real private bugs: 3-bit costs thinking time, not score — FeydRowan · 2026-11-02
- Build products for AI, not humans: MCP lets Claude Code run everything — gefei55 · 2026-11-02
- Solo dev launches hosted MCP server with 31 tools scheduling to 20+ social networks — troyizzle · 2026-11-02
- Open-source Clay killer treg promises 85% cheaper lead enrichment, runs in Claude Code — ayushtweetshere · 2026-11-02