GPT-6 Astra demos: 3D modeling, software control, and 95% zero-shot robot arm success

机器之心 · wechat · 2026-09-06

Jiqizhixin rounds up the viral GPT-6 Astra demos and their limits.

3D modeling: Astra generates runnable 3D trains via TypeScript + Three.js, and a camera exploded into 122 component groups / 1,877 parts (3h); a Tesla broke into 334 officially-sourced parts; a full human model hit 2,234 parts. Outputs aren't commercially usable yet and need human bugfixing — the real significance is a new interaction mode: describe what you want, AI translates it into spatial structure.

Computer/browser use: In a Canva test Astra drove the UI through a 1-hour multi-step portrait replication, switching to computer-use mid-task; Fable 5.1 did far worse on the same prompt. An OpenAI employee had Astra find an online piano and perform "River Flows in You" end-to-end via mouse+keyboard, then screen-record and edit the video itself. Cases like rebuilding the Palace of Fine Arts in Blender and 600 agents in one Unreal world show agents turning prompts into engineering files.

Robotics: Jay Chooi's zero-shot block-into-bowl test on a real arm: 95% success (20 runs) vs Fable 5.1's 40%, with 16% of the tokens and 43% of the cost, no demos or fine-tuning; the model outputs end-effector poses for an IK solver. He expects real-time LLM arm control by end of this year at the earliest, 2029 at the latest. Ex-NVIDIA researcher Yu Xiang argues the next frontier is truly connecting advanced models to the physical world, not just orchestrating perception/planning modules.

Takeaway: Astra remains unstable, but complex human-operated tasks are for the first time plausibly packagable into AI — skills still need learning, but what's worth learning is changing.

Original post →

GPT-6 Astra demos: 3D modeling, software control, and 95% zero-shot robot arm success

More from Embodied