Google DeepMind released Gemini Robotics 2 on July 30, giving humanoid robots whole-body control for the first time. In the same post, it published its own measured success rates across three robot bodies. Those numbers, not the demo reel, are the honest picture of physical AI in 2026.
The clip that traveled fastest was 30 seconds long. An Apptronik Apollo 2 humanoid hears a single sentence — put the watering can into the green bin on the bottom shelf — and does it. It walks to a table, picks up the can, steps over to a set of shelves, bends down, and places the can where it was told. No teleoperation, no pre-scripted trajectory. For a field where most robots are still either remote-controlled or hard-coded for one repetitive motion, that is a real result, and Google DeepMind was right to lead with it when it announced Gemini Robotics 2 on July 30.
But scroll a little further down the same announcement and you hit something more unusual than a good demo: three bar charts, published by Google, showing exactly how often its own model succeeds and fails at a list of ordinary physical tasks. Companies do not normally volunteer this. It is the most interesting thing in the release, and it is the part almost nobody quoted.
Gemini Robotics 2 is not one model. It is three, and the division of labor matters for understanding what the numbers mean.
The VLA and on-device models are limited to early-access partners. So while ER 2 is something a developer can try today, the part that actually drives a humanoid is not generally available — a distinction worth holding onto when reading breathless coverage about robots arriving in homes.
Every figure that follows is DeepMind's own reported measurement from its launch post, not an independent evaluation. That caveat matters, and it cuts in an unusual direction here: a vendor grading its own homework normally inflates the score, and these scores are not flattering. DeepMind ran the same model checkpoint across three different bodies — an Apollo 2 with SharpaWave hands, an Apollo 2 with Inspire hands, and a Franka Duo with a Robotiq gripper. Here is what it reported.
Read those three groups next to each other and a shape emerges. A two-fingered industrial gripper doing a precise insertion is nearly solved at 89.6%. A five-fingered hand attempting the same class of fine work is nowhere close. And the single most revealing pair in the entire release is 92% versus 36% — the same robot, the same hand, the same lightbulb, succeeding at taking it out and failing two times in three at putting it back.
“The gap between unscrewing and screwing is the gap between the robotics we have and the robotics we've been promised.”
DeepMind says so itself, in plainer language: multi-finger dexterous manipulation remains challenging, and its robots have more to advance in movement speed. That candor is worth noting in a year when most physical-AI communication has been a highlight reel.
The intuitive assumption is that walking is the hard part. It isn't, or at least it isn't anymore. Locomotion is a well-studied control problem with decades of research behind it, and once a model can plan a route and keep a body balanced, walking across a room becomes a solved subroutine it can call.
Fine manipulation is a different category of problem. Unscrewing a bulb is a loose task — grip anywhere, rotate, and gravity plus a bit of torque does the rest. Screwing one in is a contact-rich task: the threads have to align within a fraction of a degree, the applied force has to stay inside a narrow band, and the feedback that tells a human whether it is catching or cross-threading is tactile, not visual. A vision-language-action model is, by construction, working mostly from pixels. The SharpaWave hand on Apollo 2 has five fingers and 22 degrees of freedom, which is roughly the mechanical vocabulary needed — but having the vocabulary and knowing what to say with it are different achievements.
The quieter advance is embodiment transfer. Historically, a policy trained on one robot was worthless on another; different geometry, different sensors, different joint counts. DeepMind reports that On-Device 2 adapts to drastically different bi-arm bodies — it names the Dexmate, SO101 and Trossen platforms — in hours, typically on fewer than 200 demonstrations. If that holds outside the lab, it changes the economics of robot data collection more than any single success rate does, because it means the field stops re-learning from zero for every new chassis. It is also, notably, the one claim in the release with no accompanying success-rate chart.
Two things follow from the benchmark table, and they point in opposite directions. The first is that structured, repeatable, gripper-friendly work — warehouse kitting, machine tending, parts insertion — is closer than the discourse suggests. Franka Duo at 89.6% on precise insertion is a number you can build a business case around, given a fixture, a known part, and a human to handle exceptions.
The second is that unstructured human environments are not close. A domestic robot that succeeds at picking something off the floor 45.7% of the time is, functionally, a robot that needs supervision. The 32% dustpan result is the honest ceiling for the household-chores narrative right now. Both statements are true at once, and any humanoid pitch that doesn't distinguish between them should be read carefully.
One more distinction is worth insisting on, because the industry routinely blurs it: a model being deployed on a humanoid is not the same as that humanoid doing useful work. Gemini Robotics 2 demonstrably runs a full body. Whether a full body running it produces economic output at these success rates is a separate question the release does not answer, and a 45.7% floor-pick means every attempt needs a person watching.
The practical read: bet on grippers doing structured work in the next 24 months. Treat five-fingered humanoids doing household chores as a research program, not a product roadmap.
Gemini Robotics 2 arrives with a safety layer too — a new benchmark called ASIMOV-Agentic that tests whether the reasoning model will refuse an unsafe instruction from the action model, and whether it can recognize that a task is impossible and ask a human for help. DeepMind calls ER 2 its safest robotics model to date on constraint-following and human-proximity tests. That is a self-assessment against a benchmark the same company designed, which is worth stating plainly. It is still the right thing to be measuring if robots are going to work near people, and the fact that it is being measured at all is a maturation signal.
The broader lesson from this release is methodological. When the next lab publishes a humanoid video, three questions separate progress from theater: what was the success rate over repeated trials, on which physical body, and was the same model checkpoint used across all of them? Google answered all three unprompted. Most won't.
Physical AI is in the phase where the honest numbers are more useful than the ambitious ones. Google published both. The 36% is the one to remember.
Partly. Gemini Robotics ER 2, the embodied reasoning model, is available in Google AI Studio and in private preview on the Gemini Enterprise Agent Platform. The vision-language-action model and the on-device model — the ones that actually control a robot body — are restricted to early-access partners.
DrafterDaily reads the benchmark tables so you don't have to. One clear-eyed briefing on AI, business and technology, every weekday.
Claude Fable 5.1 costs exactly what Fable 5 cost per token. The 25-to-45% saving Anthropic advertises comes from one repriced line item — cached input, now billed at 2.5% of list instead of 10%. That is a discount you only collect if you keep the agent running.
Anthropic was not breached. The malware was already on the customer's machine, and it took a session cookie rather than a password — which is why two-factor authentication did nothing and why server-side revocation is the only lever the vendor has.
OpenAI's incident report on the Hugging Face breach leads on a security failure. The remarkable part is a verbatim chain-of-thought in which a model identifies its action as unauthorised and proceeds anyway — and an escape route that was a package manager, not a superintelligence.