A robot that knows when to stop pouring. Gemini Robotics ER 2 is built to check whether a step worked

A language model that misjudges a sentence produces a bad sentence. A robot that misjudges when a container is full spills it. Physical work punishes the gap between planning an action and confirming it happened, and that gap is what Google’s newest robotics model is designed around.

Gemini Robotics Embodied Reasoning 2, or ER 2, became available in public preview on Google Cloud on October 27, through Model Garden on the Gemini platform. It is not the model that moves a robot’s joints. It is the reasoning layer above them: it watches continuous video, understands space and time, plans long sequences of steps, reads instruments, and calls lower-level tools such as vision-language-action models, navigation APIs and custom functions. It can also coordinate several robots.

The feature that matters: knowing whether a step worked

ER 2 can classify task progress as a yes or no, a set of categories, or a continuous score. Google’s example is deciding when to stop pouring. In practice this lets a robot retry the step that failed instead of restarting a whole workflow, and it lets an operator see where in a task the system thinks it is.

Google reports 91.3% accuracy in finding the critical frame in a video (the moment an action should stop or change) with a mean error of 0.96 seconds, and about four times the execution speed of many larger model categories at a fraction of the compute. Those are Google’s benchmarks. A 91.3% detector also misses the moment almost one time in eleven, which is fine for an inspection robot that can look again and unacceptable for some industrial steps. The number tells you what to design around.

Three layers, three kinds of assurance

Who decides what in a physical agent

Reasoning (ER 2)

• Plans the task
• Reads instruments
• Judges whether a step succeeded

Probabilistic: evaluate it like a model

Control (VLA models, navigation, functions)

• Moves arms and bases
• Executes one skill at a time

Tested per skill

Safety and site rules

• Speed and force limits
• Stop zones
• Interlocks

Deterministic: never delegated to the model

Google says the model has built-in safety guardrails and that the Gemini platform adds enterprise IAM and security controls. That covers who can call the model. It does not replace the safety layer of the physical installation, which in industrial settings is governed by machine and robot safety standards and certified separately. A reasoning model that decides a step failed should be able to request a retry; it should not be able to override a force limit.

Early users named by Google include Boston Dynamics for inspections with Spot, and a self-driving materials lab project with Singapore’s A*STAR. Simulation support covers MuJoCo-Warp and NVIDIA Isaac Sim, which is where most of the testing should happen first.

When the connection is not there

The same day, Google Public Sector described a proof of concept with NASA: a robot that handles payloads and could one day prepare medical equipment for astronauts. It does not use ER 2; it uses the open Pi0 vision-language-action model, fine-tuned on Google Cloud, with an initial 85% success rate on a pick-and-place task in the lab. The useful detail is the rationale. With round trips to Earth of up to 44 minutes, models must run locally, and the robot acts under the crew’s supervision, “without ever replacing human decision-making.”

Mines, offshore platforms and remote plants face a milder version of that constraint. We saw it with Google’s tuberculosis screening AI, built to work underground without a cloud, and argued that sovereign AI is what keeps working when the connection is cut. ER 2 is a cloud model in preview. Before planning a deployment, ask where it can run, what happens to a task when the connection drops mid-step, and what the robot does by default when the reasoning layer does not answer.