In This Article
Key Takeaways
- On July 30, 2026, Google DeepMind announced three models: Gemini Robotics 2 (a vision-language-action model), Gemini Robotics ER 2 (embodied reasoning and planning), and Gemini Robotics On-Device 2 (a local variant).
- Prior versions drove a humanoid's upper body for tabletop work. This one drives the whole body, including walking, crouching and bending, demonstrated on an Apptronik Apollo 2.
- DeepMind published per-task success rates rather than a single headline score. They run from 32% to 92%, and floor-level pickups sit at 45.7%.
- Only ER 2 is generally reachable: it is on the Gemini API and Google AI Studio, built on Gemini 3.5 Flash with a 128K context window. The two action models are early-access.
- A new safety benchmark, ASIMOV-Agentic, scores whether the planner refuses unsafe tool calls and asks a human for help when it is unsure.
Most robotics announcements are a video. You watch a humanoid fold a shirt, you have no idea whether that was take one or take forty, and you move on. The Gemini Robotics 2 announcement Google DeepMind posted on July 30, 2026 has the videos too. It also has a table of success rates, task by task, hand by hand. That table is the more useful document.
What shipped on July 30
Three models, layered. Gemini Robotics 2 is the vision-language-action model (VLA): it turns camera frames and a spoken instruction into motor commands. Gemini Robotics ER 2 is the embodied reasoning model that sits above it, described in the announcement as the agent that lets "robots communicate with humans, understand the physical world and plan multi-step tasks lasting several minutes." Gemini Robotics On-Device 2 is the efficient VLA meant to run on the robot itself, without a cloud round trip.
Access is uneven, and the split tells you something. ER 2 is generally available through the Gemini API and Google AI Studio, with private preview on the Gemini Enterprise Agent Platform. The two models that actually move a robot arm are restricted to early-access partners through a Trusted Tester form. The thinking layer is open; the acting layer is not.
The whole-body claim
The headline capability is stated plainly in the post: "While our previous models controlled the humanoid's upper-body to achieve table-top tasks, Gemini Robotics 2 expands physical AI into whole-body motions." In practice that means locomotion and posture join manipulation in one control problem. Given "put the watering can into the green bin in the bottom shelf," DeepMind says the Apptronik Apollo 2 "walks to the table, and picks up the watering can, takes a few steps to the shelves, and places it precisely in its destination."
That is a real jump in scope. Reaching an object on the floor is not a longer arm motion; it is a balance problem, a footfall problem and a grasp problem at once. The published numbers reflect exactly that difficulty, which is why they are worth reading closely.
The success-rate table
Success rates as published in the DeepMind announcement
| Category | Hardware | Task | Success |
|---|---|---|---|
| Whole body | Apollo, Inspire hands | Pick up from shelf | 76.3% |
| Pick up from table | 68.4% | ||
| Pick up from floor | 45.7% | ||
| Multi-finger dexterity | Apollo, SharpaWave hands | Unscrew bulb | 92% |
| Tie trash bag | 44% | ||
| Ziplock | 40% | ||
| Screw bulb | 36% | ||
| Dustpan | 32% | ||
| Gripper dexterity | Franka Duo | Precise insertion | 89.6% |
| Diverse tool kitting | 78.9% | ||
| General pick and place | 74.2% |
Three patterns fall out. Height matters: shelf 76.3%, table 68.4%, floor 45.7%, a 30-point spread on what is nominally the same verb. Simple grippers beat fingers: every Franka Duo row clears 74%, while five-finger work swings from 32% to 92%. And direction matters more than you would guess, because unscrewing a bulb succeeds 92% of the time while screwing one in succeeds 36%. Removing a thing tolerates sloppiness. Seating a thing does not.
What happens when you chain steps
The following is our arithmetic, not DeepMind's. Per-step rates are only interesting once you multiply them, because real tasks are sequences. Take the two published rates in the watering-can demo, a table pickup at 68.4% and a shelf placement at 76.3%. If the two steps were independent, the pair completes about 52% of the time. They are almost certainly not independent, so treat that as an illustration of the shape of the problem rather than a prediction.
This is the single most portable lesson in the release, and it has nothing to do with robots. Any agent that takes a hundred actions inherits the same math. It is the reason a demo can look flawless and a deployment can fail four times out of five, and it is why we keep arguing that agent evaluation has to measure the full trajectory rather than the individual call.
ER 2 is the part you can call today
If you are a developer rather than a robotics lab, ER 2 is the piece that is actually in your hands. Per its model card, it is built on Gemini 3.5 Flash, accepts interleaved text, image, video and audio up to a 128K context window, and returns up to 64K tokens of text. It specializes, in the card's own words, in "visual, spatial, and temporal understanding, tool orchestration and success detection for physical agents."
The capability worth noticing is temporal. ER 2 works over continuous video instead of still frames, which lets it judge whether a task is going well while it is still going. Google reports 57.4% accuracy on progress classification and 91.3% on moment-finding, landing within a 0.96-second mean absolute distance of the correct frame at four times the previous generation's execution speed. SiliconANGLE quoted Google engineers Steven Hansen and Peng Xu summarizing the point: "By watching continuous video feeds, robots can now track their own progress, adapt if something goes wrong, and know exactly when to move on to the next step."
Read those two numbers side by side and the honest picture appears. Finding the instant a coffee pour should stop is close to solved at 91.3%. Judging how far along a task is remains a coin flip plus a little, at 57.4%. Self-monitoring is not one capability; it is several, at very different maturities.
Adapting to a robot you own
The On-Device model carries the most practical claim in the release. DeepMind says developers "can now adapt to new bi-arm robot embodiments with just a few hours of adaptation time, typically with less than 200 examples," and that this holds "even with new embodiments with drastically different shapes, sensors and degrees of freedom." The technique is credited to motion transfer, inherited from Gemini Robotics 1.5.
Fewer than 200 demonstrations is a small number in robot learning, where data collection is the recurring bottleneck and every new chassis has historically meant starting over. If it holds outside the lab, it changes who can afford to try. It also runs locally, which matters for the same reasons disconnected deployment matters everywhere else: latency, connectivity you do not control, and data that should not leave the building.
Safety, and the model card's own limit
DeepMind introduced ASIMOV-Agentic alongside the models, described as "a new benchmark for agentic safety orchestration and uncertainty resolution." It measures whether the reasoning agent refuses unsafe tool calls from the VLA, whether it can predict that a task is impossible, and whether it asks for human help when uncertain. Google states that ER 2 outperforms ER 1.6 and other frontier models on Safety Instruction Following and Human Proximity benchmarks.
The model card then draws a line that is easy to skim past: users must exercise discretion, and the model is not to be deployed in safety-critical settings such as healthcare or transportation, or anywhere a failure could cause injury or property damage. A model marketed for physical agents ships with a prohibition on the physical settings where failure hurts. That is not a contradiction, but it is the boundary of what was released, and it belongs in any plan built on top of it.
Why it matters
Analysis, not reported fact. The most valuable thing in this release is the disclosure norm. A frontier lab published a 32% and a 45.7% next to a 92%, in its own announcement, with the hardware named. Compare that to the culture around language models, where scores cluster in the high eighties and the interesting failures live in an appendix. Per-task rates on named hardware are reproducible claims. Aggregate scores usually are not.
For anyone buying rather than building, the transferable habit is to ask vendors for the same shape of evidence: per-task, per-configuration, with the denominator. If the answer is one number, you have been handed marketing. Our guide to evaluating an accuracy claim walks through the questions that separate the two, and the reliability math above explains why a 90% that a vendor is proud of can still produce a system that never leaves the pilot stage.
The access split is the other thing to watch. Google put the planner on a public API and kept the actuator behind a partner form. That is a reasonable safety posture and also a strategic one, and it means the near-term commercial surface here is orchestration software rather than robots. If you want to build on Gemini Robotics this quarter, you are building a planner that calls tools, watches video and knows when to stop, which is a normal agent problem with an unusual sensor. That work is available now. Whole-body humanoid control is not, and the published rates suggest why.
One claim we did not verify
The success rates above are DeepMind's own, reported in its announcement. We have not seen an independent replication, the number of trials behind each percentage, or the tolerance used to score a success. Treat them as the vendor's measurements, honestly disclosed, and still the vendor's.
Test the agent, not the demo
The same reliability math applies to every multi-step AI system, robotic or not. Start with how serious teams measure it.
Read the agent evals guideSources: Gemini Robotics 2 brings whole body intelligence to robots — Google DeepMind (July 30, 2026); Introducing Gemini Robotics ER 2 — Google blog (July 30, 2026); Gemini Robotics ER 2 model card; SiliconANGLE (July 30, 2026); MarkTechPost benchmark summary (July 30, 2026). Reliability arithmetic, framing and analysis by Precision AI Academy.
Common questions
Can I use Gemini Robotics 2 to control a robot right now? Not unless you are an early-access partner. The VLA and the On-Device model are gated behind a Trusted Tester form. ER 2, the planning model, is on the Gemini API and Google AI Studio.
What robot did the demos use? The whole-body and multi-finger results are reported on an Apptronik Apollo 2 humanoid, with Inspire hands for the whole-body tasks and SharpaWave hands for the dexterity tasks. Gripper results are on a Franka Duo. The announcement also names Boston Dynamics, Dexmate, SO101 and Trossen platforms.
Is multi-robot collaboration new here? Yes. The announcement says it "enables different types of robots to communicate and work together to solve complex workflows a single robot could not do alone," with ER 2 acting as the coordinator.
How does this compare to the previous Gemini Robotics models? The earlier models controlled a humanoid's upper body for tabletop tasks. This release adds locomotion and posture, longer task horizons described as several minutes and hundreds of decisions, multi-robot coordination, and the ASIMOV-Agentic safety benchmark. For the wider Gemini family, see our complete Gemini guide and the Gemini 3.6 Flash release.