The Sequence AI of the Week #908: You Need to Learn About Gemini Robotics

There's a moment in Google's new demo where a robot walks across a room, picks up a watering can, and drops it in the right bin. It's deliberately unglamorous. But the boring surface is hiding the actual news.

Previous Gemini Robotics models were essentially torsos. Fixed base, tabletop tasks, no legs to speak of. If a robot needed to walk somewhere, that walking lived in a completely separate control system, bolted on and unrelated to whatever was handling the grasping. Boston Dynamics figured out how to make a robot walk years ago, using hand-tuned controllers and a lot of hydraulics. That's not new.

What's new is that in this version, walking and grasping come from the same policy, driven by the same language instruction. Apptronik's Apollo 2 doesn't call a separate navigation module to cross the room and then hand off to a different arm-control system to pick up the can. One model predicts the whole sequence of actions, treating locomotion as just another part of what it's choosing to do next, not a subsystem it invokes. That's a real shift in how these systems are architected, even if the demo itself looks like nothing special.

Google is shipping this as three separate models across two access tiers, which suggests they're thinking about deployment at different price points and capability levels rather than one flagship product. The company clearly wants developers building on this now, before competitors catch up on the unified-policy approach.