Screwing in a lightbulb is the demo DeepMind chose to headline. It's also, by DeepMind's own numbers, the thing Gemini Robotics 2 is worst at.
Take a bulb out with the five-fingered SharpaWave hand and it reportedly succeeds around 92% of the time, according to informal commentary. Thread the same bulb back in, with the same fingers, and the success rate reportedly collapses to roughly 36%.
That gap—reportedly unscrew at 92%, screw in at 36%—is the whole story of this launch compressed into one motion.
The safety benchmark for a 'universal brain' is largely a benchmark of when the brain should stop and ask a human for help.
The slogan is "One brain. For any robot." It's the line Apptronik led with when Apollo 2 stepped into the frame, and it's a genuinely seductive pitch: a single intelligence layer that drops onto any body and makes it competent.
The watering-can demo sells it perfectly. Apollo 2 hears "put the watering can into the green bin in the bottom shelf," walks across the room, bends, picks it up, and places it. One prompt, whole-body coordination, feet to fingertips.
Watch that clip and cross-hardware control looks solved. That's the point of the clip.
Then you read the figures circulating in commentary, and the framing inverts. The dustpan task reportedly lands around 32%. Screwing in a bulb, reportedly 36%. These are reportedly numbers tied to the same release that gave us the tagline, though they surface only in informal commentary.
Here's what should stop you: the failure isn't in some exotic task the model was never meant to do. It's reportedly in the exact task chosen for the highlight reel.
Removing a bulb is a coarse rotational grip—turn until it's loose. Threading one in demands you find the thread, apply gentle torque, and back off the instant it catches, or you cross-thread and crack the glass. It's the difference between "grab and turn" and "grab, sense, correct, stop."
A 22-degree-of-freedom hand reportedly has more than enough mechanical range for both. The model has the fingers. What it reportedly lacks is the closed-loop finesse to know when contact means "keep going" versus "you're about to break it."
That's not a marketing problem. That's the deployment problem, and it's exactly the one DeepMind's own reliability spread exposes.
Now layer on the claim that had hardware makers nervous. The On-Device model, DeepMind says, adapts to entirely new bi-arm embodiments in a few hours—typically with fewer than 200 examples—even when the new body has drastically different shapes, sensors, and degrees of freedom.
If true and durable, that reads like a moat-killer. It says the intelligence is portable, the body is a commodity, and whoever writes the brain owns the value.
But run the bulb numbers back through that claim. If a fully trained, hero-demo task on the flagship hand lands at 36%, what does "adapts in a few hours with 200 examples" actually deliver on a novel body doing a task nobody staged?
The honest answer: a starting point, not a product. Fast adaptation to a new embodiment is a research result about how little data motion transfer needs. It is not evidence that the adapted system clears the reliability bar any real deployment requires.
And this is where the power balance quietly shifts back toward hardware. Earlier this month Apptronik launched Robot Park—a facility whose explicit purpose is to have Apollo 2 robots collect the data that trains models like Gemini Robotics 2.
Read those two facts together. The brain adapts to new bodies with a few hundred examples, and the body maker built a factory to farm exactly those examples.
The API is not the leverage point. The data-collection loop is. Whoever owns the machine that generates per-embodiment fine-tuning data—and eats the cost of collecting it, curating it, and revalidating every task on every new hand—holds more of the value than the model endpoint does.
A swappable brain shifts the balance. It does not settle it. "One brain for any robot" still requires someone to pay, per robot, per task, to drag the reliability curve up from 36% to something a business will insure.
Which brings us to the part of the release that tells you DeepMind knows all this.
Alongside the models, the company shipped ASIMOV-Agentic, a safety benchmark for "agentic safety orchestration and uncertainty resolution." Strip the phrasing and it measures something specific: can the robot refuse an unsafe action, predict when a task isn't possible, and request human intervention when it's uncertain.
Sit with that. The safety benchmark for a "universal brain" is largely a benchmark of when the brain should stop and ask a human for help.
That's the tell. Autonomy here is not measured by how often the robot completes the task. It's partly measured by how well the robot knows when to quit.
For a software agent, this asymmetry is cheap. A failed action retries. A bad output gets regenerated. The blast radius of being wrong is a wasted API call.
A robot has to survive its own mistakes. A failed screw-in doesn't retry cleanly—it cross-threads, cracks glass, drops the tool, or drives a 22-DoF hand into a shelf. The same 56-point gap that looks like a metric on a slide is, on the floor, the difference between a completed chore and a physical incident.
So the real deployment gate for a physical agent isn't the demo's best moment. It's verified state—does the robot actually know the bulb caught the thread? It's safe stopping when a person enters the workspace. It's recovery when step three of a multi-minute task fails. And it's revalidation of all of that, per embodiment, every time the brain moves to a new body.
None of those are settled by a watering-can clip. All of them are gestured at by ASIMOV-Agentic, which is to say: acknowledged as open.
There's one number worth watching that no one has published yet. If a universal brain were genuinely converging on deployment, the metric that would prove it is an intervention rate trending down over time—the robot asking humans for help less and less as it learns.
ASIMOV-Agentic frames "request human intervention" as a capability to be measured. That framing is the admission that the intervention rate is currently the thing being managed, not the thing being eliminated.
Until some maker publishes that curve bending downward, the universal brain remains what the bulb numbers say it is.
Not a deployment-ready product. A research artifact whose reliability collapses on exactly the task it chose to show you.
One brain, for any robot—as long as you don't ask it to screw the bulb back in.