Google DeepMind’s Gemini Robotics 2 Gives Robots Whole-Body Control, but Success Rates Still Swing From 32% to 92%

Google DeepMind unveiled Gemini Robotics 2, an AI model that, for the first time in this series, gives a robot control over its entire body at once, from feet to fingertips, instead of just arms and hands. Demonstrated on Apptronik's humanoid Apollo 2, the robot walks, crouches, and reaches for objects on its own, reacting in real time to spoken instructions. Google also published hard numbers from its own testing, and those numbers, not the demo footage, are what actually show where this technology stands.

What was actually announced

In the demo footage, Apollo 2 gets an instruction like "put the watering can in the green bin on the bottom shelf," which requires walking across a room, crouching, and executing a precise grip all at once, not just reaching out an arm. The robot cleans up trash, slots a cassette into a boombox, and ties a trash bag, planning its next steps on the fly rather than following a fixed script.

The numbers Google actually published

These numbers, published directly by Google itself, are more interesting than the footage. Unscrewing a lightbulb succeeds 92% of the time, but screwing that same bulb back in succeeds only 36% of the time. Unscrewing is essentially one grip and a single-direction turn; screwing it back in requires catching the thread and holding steady pressure through the whole motion, exactly the kind of precision robots still fail at more often than they pull off. The pattern holds with the simpler Franka Duo gripper too: precise part insertion succeeds 89.6% of the time, but that task runs on a closed, repeatable motion path, unlike the open home environment Apollo 2 was tested in.

"Solving AGI in the physical world"

Google frames this directly as a push toward "general-purpose physical AI" and "solving AGI in the physical world," not just another improvement to a robot's grip. That's an ambitious claim next to success rates in the 32-44% range for tasks like using a dustpan, closing a ziplock bag, or tying a trash bag, all things a person does without a second thought. The gap between the stated goal and the current success rate isn't proof the demo was faked; it shows where the real boundary sits today between language reasoning, where Gemini models already perform very well, and physical dexterity, which still takes many attempts for every one that lands.

A safety system built alongside the model, not after it

Alongside the models, Google introduced ASIMOV-Agentic, a new benchmark that checks whether a robot refuses an unsafe instruction, correctly judges whether a task is even feasible, and asks a human for help when uncertain. On top of that, the system now detects human proximity and automatically stops the robot when someone gets too close. Google calls this its "safest robotics model to date in safety constraint following and human proximity benchmarks." Building that kind of framework alongside the model itself, rather than in response to a later incident, suggests the company treats the growing autonomy of these robots as a real risk, not just a line in a press release.

Google isn't building its own robot

It's worth being clear about who Google isn't in this race: a robot manufacturer. Tesla started mass production of Optimus in January 2026 and already has tens of thousands of units out; Figure AI has its Figure 03 running on BMW's Spartanburg production line, billed by the robot-hour; and Chinese companies control most of the global market for humanoid components and finished platforms today. Instead of competing on sheet metal and motors, Google is selling the brain: Gemini Robotics 2 runs on Apptronik's hardware, but the model's architecture is built to work across many different robots at once, from industrial arms to full humanoids. It's a bet that looks a lot like the one Google made with Android: you don't need to win the hardware war if your model runs on everyone else's hardware.

SPAWNSY verdict