Assess VLM Robot Control
Yes, I think there's a meaningful chance that within 3–6 months, frontier VLMs become capable of general-purpose, low-frequency robot manipulation across a broad range of tasks. But I would distinguish two hypotheses: General control is mostly solved at the intelligence level. A sufficiently powerful VLM understands what to do and can generate reasonable actions on unfamiliar tasks, given enough inference time. The remaining problems are compression and reliability. We need to make this intelligence fast, cheap, precise, and consistent enough for real-world deployment. I'm increasingly sympathetic to both. But I think hypothesis 2 is incomplete: the ability to understand and approximately execute a task is not necessarily sufficient to learn robust, contact-rich motor control. Evaluation Astra performance Interpretation The particularly striking result is NVIDIA's October 14 Astranomy study. Astra achieved 74.8% on RoboLab-120 in its agentic configuration, but direct inference averaged