Source: Unite.AI
Google has launched Gemini Robotics ER 2, an embodied-reasoning model built to serve as a robot’s high-level planner while a separate model handles the motors. It is available to developers through the Gemini API and Google AI Studio, with private-preview access on the Gemini Enterprise Agent Platform.
ER 2 takes in video, images, audio and text, reasons about the scene, plans a task that runs several minutes, then calls something else to move the hardware: a vision-language-action model, a navigation API, or a developer’s own function declared to the model as a tool. It can also call Google Search mid-task. The model card describes it as a vision-language model based on Gemini 3.5 Flash, with a 128,000-token context window and up to 64,000 tokens of output.
It succeeds Gemini Robotics-ER 1.6, released on April 14, 2026, which added instrument reading (interpreting a pressure gauge or a chemical sight glass), a capability Google developed with Boston Dynamics for facility inspection rounds. ER 2 extends that to digital displays, linear scales, rulers and liquid thermometers, tested across 10 instrument types.
What continuous video adds
The substantive change is that ER 2 reasons over a live video feed rather than still frames, which lets it judge something that has quietly limited robot autonomy: whether a step is finished.
Two capabilities come out of that. Progress classification asks the model to sort each frame of a feed into one of five completion bands, from 0–20% through 80–100%; Google reports 57.4% accuracy. Moment finding asks it to pin the exact frame where a critical event occurs, such as the point at which a cup is full enough to stop pouring. The company reports 91.3% accuracy on that task with a mean error of 0.96 seconds, and says the model delivers it at roughly four times the execution speed of much larger models for a fraction of the compute.
Sub-second timing is the part that matters on a real machine. A robot that can tell a step is 60% done, or has failed outright, can retry that step instead of restarting the workflow or waiting for an operator. ER 2 streams through the Gemini Live API’s bidirectional endpoint, which is how Google keeps the reasoning loop from inserting stop-and-think pauses between actions. Getting inference latency low enough for physical control is the same constraint that pushes robotics vendors toward dedicated on-robot silicon and single-GPU real-time models.
Two robots, one job
The other new capability is multi-robot collaboration: different machines sharing a semantic understanding of a task and handing work between them, on the reasoning that a wheeled base and a pair of legs are suited to different parts of the same 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 does not build, which is how most of this market now works: the model comes from a frontier lab and the arms, legs and grippers come from partners, the same split behind cobot makers courting foundation-model developers and platform-level embodied stacks.
ER 2 arrived as part of a three-model family. The companion research post from the Gemini Robotics team covers Gemini Robotics 2, an action model that controls full humanoids from feet to fingertips, and Gemini Robotics On-Device 2, which runs locally and adapts to a new bi-arm robot in a few hours from fewer than 200 examples. Both go to early-access partners rather than the open API.
The team also published the success rates behind that action model, which drove three different robot bodies from a single checkpoint. Apollo 2 fitted with Inspire hands picked objects from a shelf 76.3% of the time, from a table 68.4% and from the floor 45.7%. Five-fingered work with the 22-degree-of-freedom SharpaWave hand is further back: 92% for unscrewing a light bulb, 36% for screwing one in, 44% for tying a trash bag. DeepMind describes multi-finger dexterous manipulation as still challenging, which sets a useful marker for anyone sizing up humanoid capability claims.
Safety limits and first deployments
Google reports gains on safety-instruction-following and human-proximity benchmarks, and says ER 2 halts a humanoid when a person comes close, then resumes on its own once the area is clear. DeepMind calls stopping for a nearby human a key requirement in collaborative safety standards, and it is a requirement usually met in hardware rather than in the planning model: Apollo 2’s own design pauses movement when an object enters its defined impact radius.
DeepMind also released ASIMOV-Agentic, a benchmark for whether a reasoning model behaves safely as an orchestrator: refusing unsafe tool calls from the action model, judging whether a task is physically feasible, and asking a human for help when it is uncertain.
The model card draws a firm deployment boundary. Google tells developers not to use the robotics models for safety-critical work, naming healthcare and transportation, or anywhere a malfunction could foreseeably cause death, injury or property damage. That language points the first wave of deployments toward inspection, warehouse handling and lab tasks.
For robot builders, ER 2 is the layer available to call from the Gemini API without a partnership, aimed at teams that already have working hardware and need something to decide what it does next.
