Gemini Robotics 2 Unscrews a Bulb 92% of the Time. A Dustpan? 32%.

TL;DR

Google DeepMind released Gemini Robotics 2 on July 30, a three-model stack for physical robots. The headline capability is real: one learned policy now drives an entire humanoid, legs and torso and arms and fingers, instead of separate subsystems handing control to each other. The headline numbers are more honest than the marketing usually allows. Picking an object off a shelf works 76.3% of the time. Off the floor, 45.7%. Unscrewing a light bulb with a 22-degree-of-freedom hand hits 92%, but sweeping with a dustpan lands at 32%. Only the reasoning model is publicly available; the two models that actually move a robot are behind an application form.

What actually shipped

Three models, three very different access tiers:

• Gemini Robotics 2, the vision-language-action (VLA) model that converts what a robot sees and hears into motor commands. Early-access partners only.
• Gemini Robotics ER 2, the embodied-reasoning vision-language model that plans multi-step jobs and orchestrates the VLA. Available now via the Gemini API and Google AI Studio, with a private preview on the Gemini Enterprise Agent Platform.
• Gemini Robotics On-Device 2, a compact VLA that runs locally with no network round trip. Trusted Testers program.

The whole-body claim is the genuinely new thing

Until now, a humanoid doing a chore was really several controllers in a trench coat. A locomotion policy handled balance and walking, a separate manipulation policy handled the arms, and a hand-tuned layer negotiated the handoff between them. Every new behavior that needed the legs and the arms to cooperate meant rebuilding that negotiation by hand.

Think of it as a relay race where the runners were excellent and the baton pass was a coin flip. Gemini Robotics 2 collapses the whole thing into a single end-to-end policy, so one runner covers the entire track.

That is why the floor number matters. Reaching for something on a shelf is mostly an arm problem. Reaching for something on the floor requires the robot to squat, shift its center of mass, keep its balance, and close a grasp, all at once, from one policy. DeepMind published all three heights on the Apptronik Apollo 2, and the difficulty gradient is exactly what you would expect.

Fingers remain the hard part

The dexterity results, run on a SharpaWave hand with 22 degrees of freedom, are where the gap between demo reel and deployment lives. Unscrewing a light bulb is a rigid object, a fixed axis, and a predictable contact patch, and the model nails it. A trash bag is a deformable, self-occluding mess with no rigid geometry to plan against, and the number falls by half.

Note the asymmetry between unscrewing a bulb (92%) and screwing one in (36%). Removal only needs a grip and a rotation. Insertion needs alignment, threading, and force control against a socket you cannot see. Anyone who has changed a fixture on a ladder while holding a phone flashlight already knows this in their body.

The two-arm gripper results on Franka Duo hardware are considerably stronger: 89.6% on precise insertion, 78.9% on diverse tool kitting, 74.2% on general pick and place. Grippers are easier than hands, which is a fine thing to remember the next time a humanoid render shows five articulated fingers doing origami.

ER 2 is the part you can use today

ER 2 is a vision-language model rather than a motor-control model, and it is the only piece open to everyone. Its job is to watch a continuous video feed, decide whether the task is progressing, notice when it has gone wrong, and reroute. Google reports 57.4% accuracy on progress classification, sorting video frames into five completion buckets from 0-20% through 80-100%, and 91.3% accuracy on moment finding with a 0.96 second mean absolute distance.

Read those two together. Random guessing across five buckets is 20%, so 57.4% is real signal but nowhere near reliable supervision. Finding when something happened is close to solved; judging how far along a task is remains genuinely hard. Google also pitches sub-second latency and roughly 4x the execution speed of larger models, which is the number that decides whether a planner can sit in a control loop at all.

If you build agents, this is the interesting handle: ER 2 orchestrates tools and other models, including the VLA, and integrates with the Gemini Live API for low-latency streaming. It is a planner with eyes, and you can call it without owning a robot.

Safety got a benchmark, not a press release

DeepMind also shipped ASIMOV-Agentic, an extension of the ASIMOV semantic-safety benchmark introduced with Gemini Robotics 1.0. The agentic version tests something specific and useful: whether the reasoning model refuses an unsafe command coming from the action model, whether it flags an impossible task, and whether it asks a human for help instead of improvising.

Google reports improvements over the previous ER model on safety instruction following and human proximity, including halting a humanoid when a person approaches. Carolina Parada, who heads robotics at DeepMind, framed the stakes plainly in Engadget's coverage: putting these systems into a wider range of physical situations makes the safety question more pressing, not less.

Adaptation is the quiet win

On-Device 2 adapts to a new bi-arm robot body with fewer than 200 demonstration examples in a few hours, across platforms including SO101, Dexmate and Trossen. The VLA has also been shown on Boston Dynamics Spot and Franka hardware. Embodiment portability at that sample count is arguably a bigger deal than any single success rate, because it decides whether one model can cover a fleet of mismatched robots or needs a bespoke run per chassis.

What to take from this

Physical AI is now on the same trajectory as language models circa 2021: the architecture question is converging on one big end-to-end policy, and the remaining gap is data and contact physics. A robot that unscrews bulbs 92% of the time and sweeps 32% of the time is not one you would hire, but it is one you would watch very closely.

What is missing is exactly what is always missing: no third-party evaluations, no published wall-clock task times, and the demo videos come with an acknowledged caveat that the robots move slowly and pause to think. The numbers Google chose to publish are more candid than the norm. They are still Google's numbers.

Key Takeaways

• Gemini Robotics 2 is the first Gemini model to drive a full humanoid, legs through fingers, under a single learned policy instead of stitched-together subsystems.
• Published whole-body pick rates on Apollo 2: 76.3% shelf, 68.4% table, 45.7% floor. Balance is what the last 30 points buy.
• Multi-finger dexterity with a 22-DOF SharpaWave hand ranges from 92% (unscrew bulb) down to 32% (dustpan). Deformable objects are still unsolved.
• Only ER 2 is publicly available, via the Gemini API and AI Studio. The VLA and On-Device 2 require an application.
• ER 2 scores 91.3% on moment finding but 57.4% on progress classification against a 20% chance baseline, so "did it happen" is far easier than "how far along is it".
• On-Device 2 adapts to a new robot body with fewer than 200 examples in a few hours, which may matter more commercially than any single benchmark.

Sources: Google DeepMind: Gemini Robotics 2 brings whole body intelligence to robots, Google: Gemini Robotics ER 2, MarkTechPost, Engadget, arXiv: Generating Robot Constitutions and Benchmarks for Semantic Safety