What's this about?
Google DeepMind announced Gemini Robotics 2, its next-generation "physical AI" model — pitched as "one brain for any robot" — bringing full-body intelligence to humanoids. A demo shows it helping Apptronik's Apollo 2 humanoid use whole-body coordination to pack a bag for a sports game.
What exactly is "physical AI / a VLA model"?
Most AI you meet lives in text or images. Physical AI (often a vision-language-action, or VLA, model) controls a body: it sees the scene, understands an instruction in plain language, and outputs the motor actions to carry it out. "One brain for any robot" means the same model can drive different robot bodies rather than needing a bespoke controller per machine — a general driver, not a custom one per car.
Why it matters
Generality is the hard part in robotics: a model that transfers across bodies and does whole-body (not just arm) manipulation is a meaningful step toward humanoids that are useful outside a lab, and it puts DeepMind squarely in the humanoid race.