GPT-6 Astra generates the best zero-shot Real2Sim assets tested against Opus 5.5 and Fable 5.1. It’s also the most expensive. We gave frontier models ~20 iPhone captures of real-world objects and asked them to recreate them as realistic, functional NVIDIA Isaac Sim assets: 🧵

All tests were zero-shot without human intervention. Inputs were 12-25 ordinary iPhone photos from different angles, plus 1-4 short videos of the objects and their mechanics. We then placed, tested and rendered the resulting assets in our own Isaac Sim environment. Even where zero-shot generation falls short, frontier models have become remarkably useful for asset generation. With a few days of iteration, we built a realistic, functional vacuum cleaner for use in our private evals.

Astra is dominating existing robot policies. But what can Astra do that existing policies could not? We analyzed 10 tasks, 5 Episodes each on Astra, Pi0.5, Nvidia’s Cosmos3 Nano and Gr00t 1.7 on RoboLab. 1) Astra is by far the most reliable model in terms of object recognition:

4) Simple pick & place tasks are saturated by all policies.