For two years, the demos that made AI-powered robots look clever were mostly a single arm on a tabletop, pushing blocks and folding cloth. On Thursday, Google DeepMind tried to move the story from the table to the whole body. Its new Gemini Robotics 2 release aims a single AI stack at controlling a humanoid from its feet to its fingertips — and at getting several robots to work together. It is a real step. It is also, for now, something almost nobody can actually use.

What was announced

DeepMind shipped three models, each doing a different job in the robot's head:

• Gemini Robotics 2 — the vision-language-action model, or VLA. This is the part that turns what a robot sees and hears directly into motor commands. It now drives full humanoids and two-armed systems, and DeepMind says it manages a dexterous, 22-degree-of-freedom five-fingered hand rather than just simple grippers.
• Gemini Robotics ER 2 — the "embodied reasoning" model, a higher-level planner that breaks a vague instruction into steps and can coordinate more than one robot.
• Gemini Robotics On-Device 2 — a lighter VLA that runs locally on the machine, for when a cloud round-trip is too slow or unavailable.

The showcase example is telling: controlling Apptronik's Apollo 2 humanoid, the system is told to "put the watering can into the green bin in the bottom shelf," and it walks to the table, picks up the can, and does it — planning, locomotion and manipulation stitched into one instruction. DeepMind also highlights fast adaptation, claiming a model can be taught a new two-armed robot body in "a few hours" and "typically with less than 200 examples," and multi-robot collaboration across different machine types.

The honest read

Three caveats keep this a milestone rather than a revolution.

Most of it is gated. Only Gemini Robotics ER 2 — the reasoning layer — is broadly reachable, via Google AI Studio and a private preview on Google's enterprise agent platform. The two action models that actually move the robot are limited to early-access partners. So the impressive whole-body control is, for the moment, a capability you read about rather than run.

The success rates are moderate, and they are DeepMind's own. The company reports whole-body manipulation success from roughly 46% to 76% across tasks, and gripper dexterity in the 74–90% range. Those are real numbers for seriously hard, open-ended physical tasks — but a robot that completes a chore half to three-quarters of the time is a research result, not a butler. As ever, the figures are the vendor's, measured on its own tasks, and await independent replication.

Hardware is the gate physics won't lift. A better brain still needs a body, and the named platforms — Apptronik's Apollo, Franka, Dexmate, Trossen and others — are expensive, scarce research machines. Software like this raises the ceiling on what those bodies can do; it does not put a humanoid in a warehouse this quarter.

Why it still matters

The interesting move is strategic. DeepMind is positioning Gemini as the brain for other people's robots, not building its own — the same platform play Google has run in phones and cloud, now pointed at embodied AI. That lands in a loaded week: days after the US added Chinese humanoids and robot-dogs to a national-security blocklist, a Western lab is making a credible bid to supply the intelligence layer for whatever robots the West does build.

DeepMind also paired the release with an ASIMOV-Agentic safety benchmark and work on human-proximity detection — an acknowledgement that a model fluent enough to walk a humanoid across a room is also one you want to be very sure about before it does. The demos are gated; the direction is not subtle. The race to build the mind for the machine's body just moved off the tabletop.

The short version:

• Google DeepMind released Gemini Robotics 2 on 30 July — three models: a whole-body VLA, an embodied-reasoning planner (ER 2), and an on-device VLA.
• It moves embodied AI from single-arm tabletop tasks to whole humanoid control, 22-DoF dexterity and multi-robot coordination, with claimed adaptation to a new robot body in "a few hours."
• Only ER 2 is broadly available (AI Studio + private preview); the action models are early-access only, and success rates (a DeepMind-reported ~46–76%) are moderate.
• The play is Gemini as the brain for other people's robots — landing days after Washington blocklisted Chinese humanoids.