Google has launched Gemini Robotics ER 2, an embodied-reasoning mannequin constructed to function a robotic’s high-level planner whereas a separate mannequin handles the motors. It’s obtainable to builders by means of the Gemini API and Google AI Studio, with private-preview entry on the Gemini Enterprise Agent Platform.
ER 2 takes in video, photographs, audio and textual content, causes concerning the scene, plans a activity that runs a number of minutes, then calls one thing else to maneuver the {hardware}: a vision-language-action mannequin, a navigation API, or a developer’s personal perform declared to the mannequin as a software. It may possibly additionally name Google Search mid-task. The mannequin card describes it as a vision-language mannequin primarily based on Gemini 3.5 Flash, with a 128,000-token context window and as much as 64,000 tokens of output.
It succeeds Gemini Robotics-ER 1.6, launched on April 14, 2026, which added instrument studying (decoding a stress gauge or a chemical sight glass), a functionality Google developed with Boston Dynamics for facility inspection rounds. ER 2 extends that to digital shows, linear scales, rulers and liquid thermometers, examined throughout 10 instrument sorts.
What steady video provides
The substantive change is that ER 2 causes over a stay video feed fairly than nonetheless frames, which lets it choose one thing that has quietly restricted robotic autonomy: whether or not a step is completed.
Two capabilities come out of that. Progress classification asks the mannequin to kind every body of a feed into one in every of 5 completion bands, from 0–20% by means of 80–100%; Google stories 57.4% accuracy. Second discovering asks it to pin the precise body the place a crucial occasion happens, corresponding to the purpose at which a cup is full sufficient to cease pouring. The corporate stories 91.3% accuracy on that activity with a imply error of 0.96 seconds, and says the mannequin delivers it at roughly 4 instances the execution velocity of a lot bigger fashions for a fraction of the compute.
Sub-second timing is the half that issues on an actual machine. A robotic that may inform a step is 60% executed, or has failed outright, can retry that step as a substitute of restarting the workflow or ready for an operator. ER 2 streams by means of the Gemini Dwell API’s bidirectional endpoint, which is how Google retains the reasoning loop from inserting stop-and-think pauses between actions. Getting inference latency low sufficient for bodily management is similar constraint that pushes robotics distributors towards devoted on-robot silicon and single-GPU real-time fashions.
Two robots, one job
The opposite new functionality is multi-robot collaboration: totally different machines sharing a semantic understanding of a activity and handing work between them, on the reasoning {that a} wheeled base and a pair of legs are suited to totally different components of the identical job. DeepMind’s demonstration pairs Apptronik’s Apollo 2 humanoid with a Franka Duo bi-arm platform. A second demo has ER 2 driving the navigation and manipulator APIs on Boston Dynamics’ Spot to fetch an object on a spoken command.
All of that runs on {hardware} Google doesn’t construct, which is how most of this market now works: the mannequin comes from a frontier lab and the arms, legs and grippers come from companions, the identical cut up behind cobot makers courting foundation-model builders and platform-level embodied stacks.
ER 2 arrived as a part of a three-model household. The companion analysis put up from the Gemini Robotics workforce covers Gemini Robotics 2, an motion mannequin that controls full humanoids from toes to fingertips, and Gemini Robotics On-Gadget 2, which runs regionally and adapts to a brand new bi-arm robotic in a couple of hours from fewer than 200 examples. Each go to early-access companions fairly than the open API.
The workforce additionally printed the success charges behind that motion mannequin, which drove three totally different robotic our bodies from a single checkpoint. Apollo 2 fitted with Encourage arms picked objects from a shelf 76.3% of the time, from a desk 68.4% and from the ground 45.7%. 5-fingered work with the 22-degree-of-freedom SharpaWave hand is additional again: 92% for unscrewing a lightweight bulb, 36% for screwing one in, 44% for tying a trash bag. DeepMind describes multi-finger dexterous manipulation as nonetheless difficult, which units a helpful marker for anybody sizing up humanoid functionality claims.
Security limits and first deployments
Google stories positive factors on safety-instruction-following and human-proximity benchmarks, and says ER 2 halts a humanoid when an individual comes shut, then resumes by itself as soon as the world is obvious. DeepMind calls stopping for a close-by human a key requirement in collaborative security requirements, and it’s a requirement normally met in {hardware} fairly than within the planning mannequin: Apollo 2’s personal design pauses motion when an object enters its outlined affect radius.
DeepMind additionally launched ASIMOV-Agentic, a benchmark for whether or not a reasoning mannequin behaves safely as an orchestrator: refusing unsafe software calls from the motion mannequin, judging whether or not a activity is bodily possible, and asking a human for assist when it’s unsure.
The mannequin card attracts a agency deployment boundary. Google tells builders to not use the robotics fashions for safety-critical work, naming healthcare and transportation, or wherever a malfunction may foreseeably trigger dying, damage or property harm. That language factors the primary wave of deployments towards inspection, warehouse dealing with and lab duties.
For robotic builders, ER 2 is the layer obtainable to name from the Gemini API with out a partnership, aimed toward groups that have already got working {hardware} and want one thing to determine what it does subsequent.
