NVIDIA's SpatialClaw uses code as action interface, beats prior agent by 11.2 points on 20 benchmarks
CMHungSteven · x · 2026-09-17
NVIDIA Research released SpatialClaw, a training-free spatial reasoning agent built on the thesis that "code is the right action interface":
- The agent writes Python in a persistent kernel, composing perception modules, inspecting intermediate results, and revising strategy across steps.
- Compared to single-pass code generation (can't revise) and structured tool-calls (limited composition), SpatialClaw beats a recent agent by +11.2 points across 20 benchmarks and improves consistently across six VLM backbones.
- No benchmark- or model-specific adaptation.
The author notes GPT-6 Astra's reported 3D abilities via code align with their findings; claims about Astra's >100k-GPU Stargate pretraining remain unconfirmed rumor.
Related event: NVIDIA Open-Sources Training-Free Spatial Reasoning Agent SpatialClaw(2 posts)→
More from Models
- Claude upgrade-page traffic rebounded to 17.6M in September, near May's 18.6M peak — firstadopter · 2026-11-02
- Open-source GLiNER-PII scrubs 13 identifiers from a discharge note in 550ms on GPU — MaziyarPanahi · 2026-11-02
- OpenAI Dots user says bloat and broken infra drove him to try newcomer Muse — Koala_Confused · 2026-11-02
- Fable 5.5 reportedly one-shots animation in pure code as Anthropic's coding bet pays off — Hesamation · 2026-11-02
- One agent goal has been running for 4 days, burning 30K tokens — CtrlAltDwayne · 2026-11-02
- Daily AI Brief: Gemini 4 Carbon spotted on Arena, DeepSeek open-sources coding agent — testingcatalog · 2026-11-02