Episode Details

When a humanoid talks, walks, and picks up a cup in one smooth interaction, it looks like a single mind at work. It isn't. Under the hood, most deployed robots run a stack: a vision language model doing the reasoning, a vision language action model handling manipulation, and separate systems for speech, balance, and facial expression — all wired together by orchestration software. This episode digs into the technical reality behind the demos, from Ludo Robotics' Ludi and Figure's three-timescale Helix architecture to Google's Gemini Robotics ER 2 and the single-model bets from Pelican Unify and ACE Brain. The field is split on whether modular is a temporary patch or the correct endgame. And the hardest unsolved problem isn't the flashy stuff — it's manipulation, where the best models still fail half the time on tasks like turning a sock inside out. Episode #892232 — open it directly at myweirdprompts.com/892232