Gemini Robotics 2 takes the whole body — and publishes the numbers where it still fails

Google DeepMind released Gemini Robotics 2 on July 30, 2026 — three models, three access tiers. The headline capability is whole-body control: where the previous generation drove only a humanoid's upper body for table-top manipulation, one checkpoint now walks Apptronik's Apollo 2 to a shelf, crouches and places an object. The instruction used in the worked example is "put the watering can into the green bin in the bottom shelf."

The more informative part of the release is the evaluation table, which Google DeepMind published in full rather than in highlights. It shows a model that is good at gross manipulation and still unreliable at the fine work that makes a humanoid economically useful.

Three models, three jobs

Gemini Robotics 2 is the vision-language-action (VLA) model — it turns camera and language input into motor commands, driving full humanoids as well as bi-arm platforms, on both multi-finger hands and parallel grippers.

Gemini Robotics ER 2 is the embodied-reasoning layer: a vision-language model that talks to humans, understands the scene and plans multi-step tasks lasting several minutes. Its model card puts it on Gemini 3.5 Flash, with interleaved text, image, video and audio input at up to 128k context and up to 64K output tokens. The architectural point is that ER 2 plans and then hands motor execution to a VLA registered as a callable tool — the same pattern as function calling, applied to actuators.

Gemini Robotics On-Device 2 is the local VLA, built on Gemini Robotics 1.5 technology and Google's on-device Gemma models, for deployments that cannot depend on connectivity.

The dexterity numbers

All reported rates come from a single checkpoint driving three different bodies — itself the harder claim, since cross-embodiment transfer has been the standing weakness of the VLA approach.

On Apollo 2 with Inspire hands, whole-body manipulation averages 76.3% picking from a shelf, 68.4% from a table and 45.7% from the floor — each figure an average across several tasks in that category. On Apollo 2 with the five-fingered, 22-degree-of-freedom SharpaWave hand, Google DeepMind reports per-task rates instead, and the spread is wider: unscrewing a lightbulb 92%, tying a trash bag 44%, sealing a ziplock 40%, screwing in a bulb 36%, dustpan work 32%. The Franka Duo with a two-finger parallel gripper is the most consistent of the three, averaging 89.6% on precise insertion tasks, 78.9% on tool kitting and 74.2% on general pick-and-place.

A 32% success rate on a dustpan is not a failure of the release. It is the first time anyone has put a number on it in public. — EW analysis

Read that as a shape, not a scoreboard. Rigid objects with a defined grasp axis are close to solved. Deformable material, bimanual coordination and tasks requiring sustained force control are not. Google DeepMind is also explicit that its robots "have more to advance in movement speed" — the demos are slow.

Knowing when a task is done

ER 2's headline additions target a problem that rarely gets benchmarked: temporal judgment. Progress classification — assigning each video frame to one of five completion bands — reaches 57.4% accuracy, which Google DeepMind says beats both its previous generation and competing frontier models. Moment finding, identifying the exact frame where a critical event occurs (when to stop pouring), reaches 91.3% accuracy with a 0.96-second mean absolute distance, at what the company describes as four times the execution speed of much larger models.

Success and failure detection now runs on raw video rather than static snapshots, which is what lets it catch a spill mid-execution instead of after. Instrument reading extends past circular dials to digital displays, linear scales, rulers and liquid thermometers across 10 instrument types — a mundane capability that matters a great deal for industrial inspection work.

Adaptation and the multi-robot piece

On-Device 2 is natively multi-embodiment and inherits motion-transfer techniques from Gemini Robotics 1.5. Google DeepMind reports adaptation to new bi-arm platforms in a few hours, typically with fewer than 200 demonstrations, across bodies with different shapes, sensors and degrees of freedom. The model card's scaling comparison is the clearest evidence: on the SO101 platform, introduced only during post-training, On-Device 1 moved from 0.0% to 6.7% while On-Device 2 went 6.7% to 53.3%. On Dexmate, 13.3%–33.3% became 24.4%–75.6%. Both sets come from the On-Device 2 model card.

That number is the one with commercial consequences. If a new arm can be brought up on fewer than 200 examples, the data moat around any single robot platform gets thinner — which cuts against the logic driving China's demonstration-data buildout, though it does not eliminate it: the same card notes On-Device 2 still generalizes poorly out of distribution and struggles with high-degree-of-freedom robots.

Gemini Robotics 2 also introduces handoffs between different robot types on shared semantic understanding, demonstrated with Apollo 2 working alongside a Franka F3 Duo, and a separate demo used ER 2 to orchestrate Boston Dynamics Spot navigation and manipulator APIs. Alongside the models, Google DeepMind published ASIMOV-Agentic, a safety benchmark, on Hugging Face — tagged CC-BY-4.0, though the repository is gated and its README text cites the MIT license.

Why it matters

Apptronik's Apollo 2 is now the reference body for the most capable publicly documented robot foundation model, which is a meaningful position for a US hardware maker in the week the FCC closed the door on new foreign humanoid imports. And the access split — reasoning layer open in preview, motor-control layer gated to partners — tells you where Google DeepMind thinks the defensible asset sits.