Google’s robot brain, Gemini Robotics ER 2, opens in preview on Google Cloud

Google has made Gemini Robotics Embodied Reasoning 2 (ER 2) available in public preview on Google Cloud, according to its announcement. The model acts as a robot’s high-level planner, and Google says it can locate a key video moment with 91.3% accuracy.

A planner that treats robot controls as tools

Google describes ER 2 as the core of Gemini Robotics 2, introduced earlier this year. It is a specialized Gemini model meant to sit above a robot’s low-level controllers. Developers declare those controllers, such as Vision-Language-Action (VLA) models, navigation APIs or custom functions, as tools. They then stream video, audio or text into the model, which decides the next step.

Google lists several capabilities:

• Task verification: the model judges whether a step worked, using binary status, categories or continuous scores, so a robot can retry one failed step instead of restarting a workflow.
• Event timing: it finds the exact frame where something happens, such as when to stop pouring a liquid. Google reports a mean absolute frame distance of 0.96 seconds, at 4x execution speed.
• Instrument reading: it reads values from more than 10 instrument types, including dials, sight glasses, digital displays, rulers and liquid thermometers.
• Multi-robot teamwork: different robot types can share an understanding of a workspace and hand off tasks.
• Natural language: operators can redirect a robot in plain speech while it explains what it is doing.

Google says the model works across embodiments, from stationary arms and wheeled rovers to quadrupeds and humanoids, and handles multi-minute tasks involving hundreds of decisions. These figures are Google’s own; the announcement does not say how they were measured, and points to a separate DeepMind post for benchmarks and a safety technical report.

How to get it and what surrounds it

The preview is reachable through Model Garden on the Gemini platform. Google points developers to documentation, pricing and sample code, including bidirectional streaming guides in the Gemini API robotics docs. The announcement text does not state prices, regions or usage limits, so teams need to check the pricing page.

The cloud pitch is a full workflow. For simulation, Google names MuJoCo-Warp and NVIDIA Isaac Sim, plus synthetic scenarios and auto-labeled workcell imagery. For training, ER 2’s pointing, moment-finding and progress classification can curate demonstration videos stored in Cloud Storage and BigQuery, and custom VLA models can then be trained through the Gemini platform, GKE or Cluster Director. For operation, workflows run with enterprise IAM and function calling, while Google Distributed Cloud handles latency-sensitive local jobs such as SLAM and path planning on factory floors, with fleet telemetry streamed into BigQuery.

Using the model to build the data that trains robots

The most interesting design choice is that ER 2 is not only a runtime brain. Google also positions it as a labeling and curation tool for the data used to train other models. That matters because of the old simulation problem: the reality gap, which our sim-to-real explainer defines as the mismatch between a simulation and the real world, caused by approximated friction, motor delays and contact forces. A model that can verify whether a task step succeeded, in real video, gives teams a way to score both simulated and real runs with the same judge.

It also fits a pattern in Google’s recent releases. Google’s general-purpose Gemini 4 Argon was opened to developers on October 16 at an introductory $2 per million input tokens and $10 per million output tokens, and Google DeepMind’s robotics post for that model described a cornhole skill learned from video through a video-to-sim-to-real pipeline. ER 2 is the robotics-specific counterpart, tuned for spatial and temporal video reasoning rather than general chat. The announcement does not give ER 2’s token pricing, so the two cannot be compared directly.

The hard part, in plain terms, is closing the loop. A robot that plans without checking its own results drifts; one that checks every step with a slow model stalls. Google’s claim of 4x execution speed and lower compute than many larger models is aimed at that trade-off, though the comparison models are not named in the announcement.

Customers describe inspection and lab work

Four organizations are quoted. Boston Dynamics’ Marco da Silva said the ability to apply AI models to video “opens a whole new world of possibilities” for Spot inspections, in connection with its AIVI-Learning product. DeepReach AI says ER 2 gives it task understanding of every recording in its data collection. Arm’s Barrie Mullins called the model “an important step” for physical AI and said Arm is exploring its effect on robotics compute platforms. ASTAR IMRE, a Singapore materials research institute, wants robots that can read a touchscreen on a legacy reactor with no API and load a centrifuge, moving toward self-driving labs.

These are vendor-curated statements, not results from independent tests, and none of them gives performance numbers. For now the preview lets enterprise teams try the claims on their own robots, but the announcement does not say when ER 2 will reach general availability.