Google DeepMind has shipped Gemini Robotics 2, and the framing matters more than any single spec. This isn’t one model — it’s three: a vision-language-action (VLA) model for whole-body humanoid control, Gemini Robotics ER 2 for embodied reasoning and task orchestration, and an on-device VLA that adapts to unfamiliar robot bodies in a matter of hours. Read together, they describe a company trying to become the operating system for physical AI rather than a builder of any particular machine.
From arms to whole body
The headline capability is scope. The previous generation focused on a humanoid’s upper body — the reach-and-grasp problems that dominate warehouse demos. Gemini Robotics 2 claims control from “feet to fingertips,” the whole-body coordination that separates a torso bolted to a table from something that can walk, balance, and manipulate at the same time. That jump is deceptively hard. Whole-body control means every action is a negotiation with gravity and momentum; a model that can fold laundry is not automatically a model that can stay upright while doing it.
DeepMind’s most telling claim is that a single checkpoint drives both Apptronik’s Apollo 2 humanoid and a Franka Duo arm setup. One brain, radically different bodies. If that generalization holds outside curated demos, it undercuts the assumption that every robot needs bespoke control software.
The real strategy is portability
The on-device model is the piece worth watching. Adapting to a new robot body in hours — locally, without a datacenter round-trip — targets the two things that have kept robotics fragmented: latency and hardware lock-in. Cloud inference is a non-starter when a robot has to react in milliseconds, and every roboticist knows the pain of retraining from scratch for new morphology.
Splitting reasoning (ER 2) from action (the VLA) is also a familiar, sensible pattern: a slower planner decides what to do and orchestrates multi-robot collaboration, while a fast policy handles how. It mirrors how the broader agent world separates planning from execution.
The caveats
Announcements are not deployments. “Supports whole-body motions” and “adapts in hours” are vendor claims until independent labs stress-test them on messy, unstaged tasks. Notably, only ER 2 appears to be broadly available so far — the flashiest whole-body control is the least accessible. And humanoids remain the least proven form factor commercially; a great brain doesn’t fix a body that’s expensive and fragile.
Still, the direction is clear. Google isn’t trying to win the humanoid hardware race — it’s trying to make sure that whoever does win, wins on Gemini. That’s the more durable position, and a more interesting one to watch.