On October 27, 2026 (US time), Google announced that it is making Gemini Robotics ER 2, an AI model that supports robots' situational judgment and task planning, available as a public preview on Google Cloud. The model was first announced on July 30 and made available to developers through the Gemini API and Google AI Studio, but it had remained a limited preview on the enterprise-focused Gemini Enterprise Agent Platform. With this expansion, enterprises can test an AI that handles a robot's task decisions while combining it with Google Cloud authentication and permission management.
Gemini Robotics ER 2 is a higher-level model that understands its surroundings from images, video and other inputs and decides how to proceed with a task. Concrete control, such as driving a robot's motors, is handled by the connected control system or by a separate AI model. This announcement therefore does not mean that an AI model in the cloud can now run robots autonomously on its own. Companies considering adoption need to evaluate separately the API specifications available in this Cloud version and the control systems required on the physical machine, rather than relying only on the demos shown in July.
The AI that plans tasks, and the control system that moves the machine
The "ER" in Gemini Robotics ER 2 stands for "Embodied Reasoning," the ability to make judgments that take a robot's body and physical environment into account. The model can grasp the positional relationships of objects from images and video, and can output information such as coordinates and bounding boxes around objects. It is also designed to break complex instructions into multiple tasks and to carry out a sequence of processing by calling robot control functions and generating code.
For example, given the instruction "put the apple in the bowl," it first recognizes the target apple, then works out the steps of grasping it, carrying it to the bowl and placing it. Gemini Robotics ER 2 handles planning the overall task and judging its progress, while the actual arm movements and grasping are delegated to a lower-level VLA (Vision-Language-Action) model, which links vision, language and action, or to the machine's control API. This division of roles makes it possible to envision a setup in which a shared decision-making model is used across different types of robots. However, interfaces tailored to each machine's mechanics and control method still need to be implemented.
The Cloud version outputs text and also supports structured output, which returns results in a defined format, as well as calling external functions. This allows object coordinates and the next action to be passed in a form that robot-side software can easily process. However, even if the model can decide on a work procedure, that alone does not guarantee safe operation such as grasping objects or avoiding collisions. In real-world use, coordination with lower-level control systems is essential.
Google also describes use cases in which the model judges progress from footage of the work in progress, detects failed steps and triggers a retry. It can also read a variety of instruments, including circular dials, liquid-level gauges and digital displays. These capabilities are intended for use in factories and research facilities, but being able to use individual functions is not the same as reliably completing an entire task on site.
What do a 91.3% accuracy rate and a 0.96-second error tell us?
Google's announcement of Gemini Robotics ER 2 in July presented results from several evaluations of video understanding. One is a task called "Moment Finding," which involves identifying a specified key moment in a video. For example, in a task of pouring liquid into a container, it evaluates the ability to find in the footage the moment when pouring should stop. On this task, Gemini Robotics ER 2 achieved an accuracy rate of 91.3%, with a mean absolute error of 0.96 seconds between the identified time and the correct one.
It is worth noting that 0.96 seconds is not the model's response time. It indicates how far, on average, the moment identified in the video deviated from the correct point in time, and does not represent the time required for inference. Likewise, the 91.3% accuracy rate is not the proportion of times a robot succeeded in a physical task from start to finish. It is simply an evaluation result for the task of finding a specific moment in footage.
Google also evaluated "Progress Classification," which judges the progress of a task as a whole. In this task, progress was classified into five bands (0–20%, 20–40%, 40–60%, 60–80% and 80–100%), and the evaluation checked whether the model could judge it correctly. Gemini Robotics ER 2 scored 57.4% accuracy, a different result from that for identifying key moments. This shows that the ability to find the point at which an action occurs and the ability to judge in fine detail how far the whole task has progressed are different challenges.
These figures are all evaluation results Google published in July and were not newly achieved with the October Cloud release. In addition, the published materials alone do not make it possible to fully grasp the scale of the evaluation data or how often misjudgments occur in real factories and similar settings. The figures of 91.3% and 57.4% therefore cannot be treated as on-site task success rates.
Cloud API specifications and changes since the July announcement
What matters most in this announcement is less any new feature of the model itself than the wider availability on an enterprise cloud platform. At the time of the July 30 announcement, developers could use the model through the Gemini API and Google AI Studio, while availability on the Gemini Enterprise Agent Platform was a limited preview. According to the Google Cloud model specifications, the Cloud version's release date is October 27, and a Google Cloud blog post of the same date announced the move to public preview.
However, not every feature introduced in the July model announcement is necessarily available in the Cloud version. Particular attention is needed on support for the Gemini Live API, which handles bidirectional streaming. The July announcement described real-time two-way communication through the Gemini API, but the Cloud version does not support the Gemini Live API.
The table below compares the July announcement with the Google Cloud model specifications dated October 27.
| Item | July 30, 2026 model announcement | Cloud version (released October 27) |
|---|---|---|
| Access route | Released to developers via the Gemini API and Google AI Studio | Available from Google Cloud's Model Garden |
| Availability on enterprise platform | Limited preview on Gemini Enterprise Agent Platform | Public preview on the same platform |
| API and bidirectional streaming | Bidirectional streaming via the Live API through the Gemini API was introduced | Supports the GenerateContent API. The Gemini Live API is not supported |
| Output and machine control | Passes high-level task decisions to lower-level VLA models and control APIs | Connects with external systems via text output, structured output, function calling and more |
Being able to input video is a different matter from being able to communicate continuously in both directions using the Live API. With this Cloud version, setups that depend on the Gemini Live API cannot be migrated as they are. At the same time, the fact that it is currently listed as unsupported does not rule out support in the future.
The Cloud version's model ID is "gemini-robotics-er-2-preview," and it accepts text, images, audio and video as input. Output is limited to text, but it supports structured output, function calling and the Count Tokens API. On the other hand, it does not support the Interactions API or the Chat Completions API, and the only region currently available is "global." Tuning, which allows additional training or adjustment of the model, is also not supported.
Enterprises can connect to the model through Application Default Credentials (ADC) using their Google Cloud credentials. Google envisions operation that uses Model Garden as the entry point, combined with permission management through Identity and Access Management (IAM) and function calling. When using the SDK, enterprise connection settings and specification of the global region are also required. However, even if cloud-side authentication and permission management are in place, developers must design for themselves the mechanism that safely passes the model's decisions to the robot's control software.
Real-machine testing is essential before deploying in factories or labs
The development process Google outlines envisions validating tasks in a simulation environment before running an actual robot. For example, motions can be checked in advance with MuJoCo-Warp or NVIDIA Isaac Sim, and training data can be prepared from generated scenarios and demonstration videos. Google says Gemini Robotics ER 2 can be used for data selection and labeling, such as extracting key scenes from videos stored in Cloud Storage or BigQuery and classifying task progress. The prepared data can then be used to train a separate custom VLA model. However, this is a mechanism for training another model, not a function for additionally training the Cloud version of Gemini Robotics ER 2 itself.
The operational stage also envisions a configuration that divides processing between the cloud and the site. Google describes placing latency-sensitive processing, such as SLAM (simultaneous localization and mapping) and path planning, on-site using Google Distributed Cloud. Meanwhile, operational information and telemetry data from multiple robots can be aggregated in BigQuery and used to analyze operating conditions. This is not an explanation that Gemini Robotics ER 2 itself runs inside the factory, but an example configuration for appropriately dividing cloud-based decision-making from on-site control processing.
Google's announcement also includes examples of use by companies and research institutions. DeepReach described using the model to understand recorded work footage and efficiently select data useful for robot training. A*STAR IMRE, a research institute in Singapore, expressed expectations for using it to read the touchscreen of an older microwave reactor that has no dedicated API, and for applications in laboratories that run experiments autonomously. However, these are usage methods and future plans described by the companies and institutions, and it has not been confirmed that large-scale commercial operation or fully autonomous laboratories have been realized.
In real deployments, response latency as well as judgment accuracy will be an important evaluation criterion. Google's developer guide explains that response times may become longer when complex queries or high-resolution inputs are used, or when the thinking level is set to "HIGH." The model may also generate incorrect information in response to ambiguous instructions or inputs that fall outside the distribution of its training data. The guide recommends cropping and enlarging the relevant part of an image when recognizing small objects, and performing multiple checks in situations that demand high accuracy.
The Gemini Robotics ER 2 now offered is in a preview stage ahead of general availability. Under Google Cloud's terms, use in production environments and for commercial purposes is permitted subject to certain obligations and restrictions, but the service is provided "as is" and support may be limited. The timing of general availability has not been disclosed at this point.
Google states clearly in its developer guide that responsibility for operating robots in a safe environment lies with the user. The Cloud release gives enterprises more options for building advanced AI into robots' situational judgment and task planning. But whether an actual machine can be operated safely and reliably depends on judgment accuracy with on-site footage and instructions, response latency, coordination with lower-level control systems, and verification of independent safety measures. The focus going forward will be how reliable this higher-level model, now available in the cloud, proves to be in actual factories and research facilities.