Gemini Robotics 2: what Google's own benchmarks say

The clip everyone shared shows a humanoid crouching to pick a watering can off the floor, walking it across the room, and setting it on a low shelf. Smooth. Slightly unsettling. What almost nobody scrolled to is the chart further down Google's own post, where picking an object off the floor is logged at 45.7 percent. That's the launch in one sentence: DeepMind announced Gemini Robotics 2 on 30 July 2026, three models covering whole-body control, planning and on-device inference, and it published the unflattering success rates right next to the highlight reel. We think that's the most interesting thing about the release.

Three models, and only one you can call

The naming is doing some work here, so let’s separate it out.

Gemini Robotics 2 is the vision-language-action model. Camera in, language in, motor commands out. The headline change from the previous generation is that it drives a whole body rather than a pair of arms bolted to a table. DeepMind’s phrasing is feet to fingertips, and the watering can demo is the illustration: the robot crouches, grasps, stands, walks, places.

Gemini Robotics ER 2 is the embodied reasoning layer. It plans the multi-step version of the task, describes what it’s doing to a person, and this generation coordinates more than one robot on the same job.

Gemini Robotics On-Device 2 is the small one, meant to run locally on the machine instead of round-tripping to a datacentre. DeepMind says it adapts to a new bi-arm embodiment in a few hours, with fewer than 200 examples. Honestly that number impressed us more than any of the demos.

Now the part that changes what you can do this week. ER 2 is in Google AI Studio, plus a private preview on the Gemini Enterprise Agent Platform. The VLA is private preview with a waitlist form. On-Device 2 is trusted testers. So the piece you can put your hands on is the planner, and the piece that actually moves a joint is a form you fill in and wait.

Read the table, not the reel

DeepMind published per-task success rates. Give them credit for that, because plenty of robotics launches ship a video and nothing else.

On an Apptronik Apollo with Inspire hands, doing whole-body manipulation: 76.3 percent picking from a shelf, 68.4 from a table, 45.7 from the floor. That ordering makes physical sense. A shelf puts the object near chest height with the body already stable. The floor means crouching, which moves the centre of mass while you’re trying to close a hand precisely, and the score drops by thirty points.

Then the fine motor set, Apollo with SharpaWave hands. Unscrewing a light bulb: 92 percent. Screwing one in: 36.

Same bulb. Same hands. Same model.

That pair is the most informative thing in the whole release, and I don’t think it got enough attention. Taking a threaded thing out is a coarse, forgiving motion. Putting it back requires aligning an axis you can’t see well, then applying steady torque without cross-threading, and any wobble in the standing body propagates all the way to the fingertips. Assembly is harder than disassembly, which is a very old lesson in robotics and apparently still true when the controller is a frontier model.

The rest of that column: trash bag 44 percent, ziplock 40, dustpan 32.

The gripper results on the Franka Duo arms sit much higher: 89.6 percent on precise insertion, 78.9 on diverse tool kitting, 74.2 on general pick and place. So this isn’t a model that misunderstands the goal. Give it a rigid two-finger gripper on a fixed base and it inserts things nine times out of ten. The losses show up when you add a standing body and twenty-odd finger joints to coordinate at once.

Which is roughly what “whole body intelligence” costs, at this point in the curve.

What the demo reel is and isn’t showing

Engadget’s write-up carries the caveat that matters most, and it came from the briefing rather than the blog: the model was specifically trained to perform every task in that video. These aren’t general-purpose machines improvising in a kitchen they’ve never seen.

Watch the grid clip with that in mind and a second detail shows up. Google labels the playback speed on every panel. Most read “Autonomous 1x”, a couple say 1.5x, and one says 4x. Labelling the speed-up is the honest choice, and vendors don’t always make it. It also tells you these robots are slower than a person at the same task, which the raw footage would otherwise hide.

One panel has an arm reaching over to tick an “I am a robot” checkbox on a monitor. Cute. Also a decent joke about where the CAPTCHA arms race has landed.

The safety material, and how much to weight it

DeepMind introduced a benchmark called ASIMOV-Agentic with this release. The previous ASIMOV work scored whether a model would refuse a harmful physical instruction. The agentic version targets orchestration: a model that plans, delegates and coordinates has more room to arrive somewhere harmful through a sequence of individually reasonable steps.

DeepMind also says ER 2 is its best model yet at following safety constraints, that it recognises when people are nearby, and that it can call a safety tool to bring a robot to a stop when someone gets too close. Carolina Parada, who heads robotics at DeepMind, framed the current phase as needing to understand the safety question more deeply, which reads to us like an honest hedge rather than a claim of solved.

Worth keeping the frame straight though. This is a vendor reporting its own scores on a benchmark it wrote. That’s normal and it’s still useful, because a published benchmark is something outsiders can eventually run. It just isn’t independent verification, and nothing here has been through one.

One loose thread: at least one write-up states that ER 2 is built on Gemini 3.5 Flash. DeepMind’s own post doesn’t say that, and the model page doesn’t either. It’s plausible, it fits the latency story, and we’re treating it as unconfirmed until Google writes it down.

So what does this change

If you build robots, not much this week, unless your waitlist form gets picked up. The reachable piece is ER 2 for planning and spatial reasoning, which is genuinely worth an afternoon in AI Studio if you’re doing anything with cameras and physical space.

If you’re trying to judge where humanoids actually are, this release is more useful than most, precisely because the numbers are in it. A machine that puts a bulb in a socket a third of the time is not ready to do that job. A machine that inserts parts with a gripper 89.6 percent of the time is close to being ready for a narrow one. Both facts came from the same announcement on the same day, and only one of them made the headlines.

We track this kind of gap a lot. Our read on the Tesla Optimus V3 reveal went the same way, and the FCC rule that quietly covers robot vacuums is the regulatory half of the same story. If a specific benchmark claim ever needs checking against the source, our tools won’t help you there, but reading the vendor’s own footnotes usually will.

Sources: Google DeepMind, Gemini Robotics 2 brings whole body intelligence to robots, the Gemini Robotics 2 model page, Engadget and MarkTechPost, 30 July 2026. Every success rate quoted here comes from DeepMind’s own published tables. The caveat that the model was trained on the tasks shown in the video, and the Carolina Parada quote, are from Engadget’s briefing. The claim that ER 2 is built on Gemini 3.5 Flash appears in MarkTechPost only and is not stated by Google, so we have marked it unconfirmed. The hero image is DeepMind’s own announcement graphic and the six-panel still is a frame from DeepMind’s demo video, both credited above. The demo video itself is hosted by Google.

Frequently asked questions