HomeBody robot architecture: cloud model only picks skills, all 250Hz control runs locally

joemeno · x · 2026-09-28

A thread on the HomeBody humanoid project argues the interesting part isn't "GPT Astra controls a humanoid" but what they refused to let the frontier model do: Astra only emits high-level skill choices like pick(target), navigate(goal) or opendrawer(handle). Once a skill starts, segmentation, stereo depth, IK, collision checking, visual servoing and bounded retries all run locally, with arm commands at 250Hz and locomotion at 50Hz — Astra is only re-consulted on success or exhausted local recovery. The author calls this layering exactly right: LLMs/VLMs are great at deciding what happens next but a ridiculous place for a 250Hz control loop. The project page's failure modes are nearly more useful than the demo: cloud reasoning adds pauses, finger servos overheat on long tasks, Real2Sim adds setup/API cost, and the local stack still needs an RTX 4090 laptop.

Original post →

More from Embodied

Embodied channel →