Google has opened public preview of Gemini Robotics ER 2, the model DeepMind built to act as a robot’s planner, to business customers through its cloud platform. Companies can find it in Model Garden, the model catalogue inside the Gemini platform, and Google pitches it as the part of a robot that decides what to do next while other software moves the joints.
The model takes in video, audio, or text and plans multi-step jobs. Google’s Cloud blog, written by Carolina Parada and Adrian Jones, says it works across robot bodies, from fixed arms and wheeled rovers to four-legged machines and humanoids. A developer registers the robot’s existing controls (vision-language-action models, navigation calls, custom functions) as tools, and the model calls them in sequence. Operators can give instructions in plain language and redirect the robot mid-task.
The most concrete claim is about checking work. A robot that cannot tell whether it dropped a part or finished a pour will keep going and compound the error. According to Google, ER 2 reads a continuous video feed and rates progress as pass or fail, as a category, or as a score, so a failed step can be retried without restarting the whole job. On finding the exact frame where something happens, such as the moment to stop pouring a liquid, Google reports 91.3 percent accuracy with a mean error of 0.96 seconds. It says the model runs at four times the execution speed of many larger models, at a fraction of their compute cost. Those numbers come from Google’s own announcement, and the Cloud post names no outside benchmark or comparison model.
Instrument reading is the quieter feature. The model can take values off more than 10 kinds of gauge, including dials, sight glasses, digital displays, rulers, and liquid thermometers, using only a camera. That matters for older plants and labs where equipment has no digital connection to read from.
Google also wires the model into the rest of its cloud. Teams can test robot behavior in MuJoCo-Warp or NVIDIA Isaac Sim before touching hardware, then use ER 2 to label demonstration videos stored in Cloud Storage and BigQuery. Google also offers three routes for training custom vision-language-action models: Google’s own AI platform, a Kubernetes service, or its cluster scheduler. For deployment, latency-sensitive jobs such as mapping and path planning move to the factory floor on Google Distributed Cloud, while fleet telemetry streams back into BigQuery. The model itself sits behind enterprise access controls and function calling.
Four customers are quoted in the post. Marco da Silva, who runs Spot at Boston Dynamics, says video changes what an inspection can be, because the robot is no longer judging a single moment. Chris Liu of DeepReach, a startup that records how skilled people work, says the model helps his company find the most useful recordings among its captured data of skilled workers. Barrie Mullins of Arm talks about exploring future robotics compute platforms. Prof. Loh Xian Jun, who runs a materials research institute at Singapore’s ASTAR, describes robots that read a touchscreen on an old microwave reactor with no API, load a centrifuge, and move samples through a busy lab. All four are Google partners, and none gives a measured result.
The post also leaves out pricing, which it routes to a separate documentation page, and it gives no date for general availability.
For anyone building robots, the choice Google is forcing is where the planning layer lives. A robotics team already on Google Cloud can now keep simulation, labeling, training, and the reasoning model under one contract, while a team on another cloud has to weigh that convenience against sending live video to Google’s models.
Reported by the Google Cloud blog (Carolina Parada and Adrian Jones); the published post carries no date.
