Gemini Robotics ER 2 Comes to Google Cloud

Google has opened a public preview of Gemini Robotics ER 2 on Google Cloud, giving businesses the robot reasoning model it released to developers in July. Google calls the model a robot’s “high-level brain”: it plans multi-step tasks, watches video to check whether each step worked, and hands the physical movements to other control software.

The model is now available “in public preview on the Gemini platform” through Model Garden, according to a Google Cloud blog post on 27 October by Carolina Parada, who heads Google DeepMind’s robotics work, and Adrian Jones. Google Cloud’s documentation lists it as a preview offering on its Gemini Enterprise Agent Platform, updated the same day.

In brief

Businesses can now use Gemini Robotics ER 2 in preview on Google Cloud, with Google Cloud security and access controls.
Google says the model finds key moments in video with 91.3% accuracy. That is Google’s own benchmark.
Promotional pricing of $1 per million input tokens runs to 31 December. The listed price from 1 January 2027 is $2.

What Google is offering

Gemini Robotics ER 2 is a vision-language model built on Gemini 3.5 Flash, according to Google Cloud’s documentation. It takes in text, images, video and audio, and returns text. Developers can declare robot controls, such as vision-language-action (VLA) models or navigation tools, as tools the model can call. The model then breaks a request such as “put the apple in the bowl” into steps and calls those tools in turn.

Google says the model works across many kinds of robot, “from stationary arms and wheeled rovers to quadrupeds and complex forms”. It also says the model lets several robots share an understanding of a workspace, split up work and hand tasks to each other.

The cloud version adds Google Cloud tooling around the model. Google says teams can test robot behaviour in simulators such as MuJoCo-Warp and NVIDIA Isaac Sim, use the model to label demonstration videos stored in Cloud Storage and BigQuery, and train their own VLA models. Local, time-critical jobs such as mapping and path planning can run on factory floors using Google Distributed Cloud, the post says.

The documentation says that during the preview the model is only served from Google’s “global” endpoint. It also says customers “may elect to use it for production or commercial purposes”.

What Google claims the model can do

Google’s headline claim is about video. It says the model can find the exact frame where a key event happens, such as when to stop pouring a liquid, “with 91.3% accuracy and a 0.96-second mean absolute frame distance”. Google adds that it runs “at 4x the execution speed and a fraction of the compute cost of many larger model categories”, but does not name the models it compared against. Google first published these figures when it launched the model to developers on 30 July. They come from Google’s own tests and have not been independently checked.

Google also says the model can read 10 or more kinds of instrument, including dials, sight glasses, digital displays, rulers and liquid thermometers. In July, Google said the model scored 57.4% on its own test of tracking how far a task has progressed.

Who is using it

Google named four organisations. Marco da Silva, who runs Spot at Boston Dynamics, said the model “brings them even more value” for customers using AIVI-Learning, the video inspection feature that TSN covered when Boston Dynamics released Spot 5.2. Chris Liu, co-founder of DeepReach, said it gives his company “accurate task understanding of every recording”.

Barrie Mullins of Arm said Arm is “excited to be working with Google on exploring” how such models could shape future robotics chips. Prof. Loh Xian Jun of Singapore’s ASTAR Institute of Materials Research and Engineering said ASTAR is using the model to give its lab robots the flexibility to work with existing equipment, such as reading the touchscreen on an old microwave reactor that has no API, or loading a centrifuge. “This brings us much closer to self-driving labs,” he said, adding: “This is the next level that we aspire to reach.”

Why video checking is the real test

Planning is only half of a robot’s job. The harder part is knowing whether a step actually worked before starting the next one. Google’s pitch for ER 2 rests on that: a model that watches a live video feed, judges whether a task has succeeded or how far it has got, and retries a failed step without restarting the whole job. Google’s named customers describe exactly those jobs: video inspections by Spot, labelling recordings of skilled work, and lab automation. TSN has covered other inspection robots this month, including ANYbotics’ ANYmal X for explosive gas zones and Percepto’s Axion inspection software.

The evidence so far is thin, though. The 91.3% figure is Google’s own benchmark, published in July and repeated now, and the “4x” speed claim names no rival. The customer quotes are mostly about plans: Arm is “exploring”, and A*STAR calls self-driving labs a level it aspires to reach. None of the four gives a deployment figure.

Google’s documentation is also blunt about limits. It warns that the model can “hallucinate”, that complex queries can slow it down, and that “physical robots can cause damage”. It puts safety on the customer: “it is your responsibility to maintain a safe environment around the robot.” Its privacy notice goes further. Because the model uses video and audio, customers agree not to let identifiable people interact with the robots, or be near them, until those people have been told about and consented to the use of their data. The notice also asks customers to use techniques such as face blurring, and to run the robots away from identifiable people where they can. Any site with staff working near the robots will have to plan for that.

Price is the other test. Google lists promotional rates of $1 per million input tokens and $5 per million output tokens until 31 December, delivered as 50% credits back. From 1 January 2027 the listed rates are $2 and $10, so firms that start testing now will see the listed rate double after about nine weeks. Google also tied access to price tiers when it released Gemini 4 Argon earlier this month.

What happens next

The model remains a preview, which under Google’s terms means it is offered “as is” and may have limited support. Google has not said when it will be generally available on Google Cloud. The promotional pricing ends on 31 December 2026. Google DeepMind’s model page and model card still list the enterprise platform as a private preview, while the Google Cloud blog and documentation describe the new preview.