University of Maryland and AWS evaluate GPT-6 Astra for 3D scene coding, hit 53% indoor accuracy

Six GPT configurations were tested. All delivered a usable scene, but geometric accuracy varied widely. GPT‑6 Astra, the strongest model, achieved 53.4 % accuracy on indoor scenes and 39.6 % on outdoor scenes, while weaker setups hovered around 15 %. Raising the reasoning budget boosted Astra’s office subset score from 32.3 % to 61.8 %. The authors found agents struggled with self‑assessment, often mis‑ranking revisions. Their LEGO‑Plugin, which replaces self‑judgment with concrete measurements, lifted weaker models by up to 62.7 % and gave the top model a modest two‑point gain.

The generated scenes can be queried for vision tasks, yielding roughly half the object‑detection performance of the specialized DINO model, but larger gaps for segmentation and depth compared with SAM 3 and Depth Anything 3. Researchers see GPT‑6 Astra’s spatial understanding as a notable step, while industry players like Unity are already releasing plugins for agents such as Claude Code and Codex.