Gemini Robotics: what Google DeepMind has released, which robots run it, and its limits

Gemini Robotics is a family of AI models from Google DeepMind that turn camera images and spoken or written instructions into robot actions. It was announced in March 2025, and an offline on-device version followed in June 2025. DeepMind has shown it on lab arms (ALOHA 2, Franka) and on Apptronik's Apollo humanoid. Boston Dynamics is now researching it for Atlas. Most versions are still only available to selected testers.

This article sets out what DeepMind has actually published, in order, and what it means if you are watching robots move from labs towards real venues.

The first release: two models, March 2025

Google DeepMind announced Gemini Robotics on 12 March 2025 as two models built on Gemini 2.0:

• Gemini Robotics, a vision-language-action (VLA) model. It takes images and instructions and produces motor commands that control a robot directly.
• Gemini Robotics-ER (embodied reasoning), a model with stronger spatial understanding. It does not move the robot itself; roboticists use its output, such as object locations or planned grasps, in their own control software.

DeepMind presented three goals: generality (coping with objects and settings it has not seen), interactivity (following conversational instructions and reacting when things change) and dexterity (multi-step tasks such as folding origami or packing a bag). The company said the model more than doubled performance on its own generalisation benchmark compared with other VLA models. That figure comes from DeepMind's own testing, not an independent evaluation.

The technical report on arXiv, first submitted on 25 March 2025, gives more detail. It says the models can be adapted to new tasks with as few as 100 demonstrations and to new robot bodies. If the term VLA is new to you, our explainer on vision-language-action models covers how they work.

Which robots and which partners

At launch, DeepMind said most of its training data came from ALOHA 2, a two-armed research platform. It also showed the model on a two-armed Franka setup used in universities, and on Apollo, the humanoid built by Austin-based Apptronik, which it named as its main hardware partner.

A group of trusted testers was also given access to Gemini Robotics-ER: Agile Robots, Agility Robotics, Boston Dynamics and Enchanted Tools. Being a tester means evaluating the model, not shipping it in a commercial product.

The largest partnership so far came on 5 January 2026, when Boston Dynamics and Google DeepMind announced a research partnership at CES to bring Gemini Robotics models to the new Atlas humanoid. Both companies described it as joint research, starting in the following months and aimed at industrial work, beginning with car manufacturing. It is a research plan, not a product on sale. Our page on Boston Dynamics Atlas covers the robot itself.

Gemini Robotics On-Device: running without the cloud

On 24 June 2025 DeepMind released Gemini Robotics On-Device, a smaller VLA model built to run on the robot's own computer. Its main advantage is that it does not need a network connection. That matters when latency counts or when connectivity is poor, which anyone who has relied on venue Wi-Fi will recognise.

DeepMind's main points about the on-device model:

• It was trained on ALOHA robots, but developers adapted it to the bi-arm Franka FR3 and to Apptronik's Apollo humanoid.
• It can be fine-tuned to a new task with as few as 50 to 100 demonstrations.
• It came with a Gemini Robotics SDK for testing in simulation and adapting the model. Access required joining a trusted tester programme.

DeepMind was open about the trade-off. The on-device model performs worse than the full cloud-based Gemini Robotics on harder, unfamiliar tasks and on complex multi-step instructions, although DeepMind says it beat the earlier on-device models it compared it with. For safety, DeepMind advised developers to test against semantic safety benchmarks and to run red-teaming exercises, with low-level controllers handling physical safety during movement.

Gemini Robotics 1.5: thinking before acting

On 25 September 2025 DeepMind announced Gemini Robotics 1.5 and Gemini Robotics-ER 1.5. The idea is a division of labour. The ER model plans a task step by step and can call tools such as Google Search. The VLA model then carries out each step and can set out its reasoning in plain language before acting.

DeepMind also reported that motions learned on one robot transferred to others (ALOHA 2, Franka and Apollo) without separate training for each. Availability split here for the first time: Gemini Robotics-ER 1.5 was opened to developers through the Gemini API in Google AI Studio, while the action model, Gemini Robotics 1.5, stayed with selected partners. DeepMind also released an updated version of ASIMOV, its benchmark for semantic safety, which it first published in March 2025.

The second generation: what the pages show now

When we checked on 31 October 2026, DeepMind's model pages listed three second-generation models. We could not find a dated launch announcement for them, so the dates below come from the developer documentation where it gives them.

• Gemini Robotics 2. The Gemini Robotics 2 page describes it as DeepMind's most advanced VLA, available in private preview with a waitlist. It names Apollo (with Inspire and Sharpa hands) and a Franka Duo setup, and mentions whole-body control such as stepping, squatting and bending. DeepMind's own published scores vary widely by task: 92% on an unscrewing task, for example, but 44% on tying bags and 32% at the low end of its multi-finger results.
• Gemini Robotics-ER 2. The Gemini API robotics documentation lists it as a preview model, updated in July 2026, alongside a streaming variant for real-time use. The previous ER 1.6 preview, dated December 2025, is marked deprecated as of 31 August 2026.
• Gemini Robotics On-Device 2. The On-Device 2 page says it is in a trusted tester phase and can adapt to new robot bodies with fewer than 200 examples and a few hours of training. It does not name specific robots.

The pattern from 2025 continues. The reasoning model is open to developers in preview, while the models that move robots are still in private or tester access.

The limits DeepMind states itself

The developer documentation is direct about risk. It says users are responsible for keeping the area around the robot safe, that AI models make mistakes and that physical robots can cause damage. It also lists hallucinations, added latency on complex queries and significant computing costs. It asks for explicit consent from identifiable people whose voice or image the robot processes, which links to the wider privacy questions in our guide on cameras, microphones and visitor privacy.

Three further caveats are worth keeping in mind:

• Results are self-reported. The success rates and benchmark comparisons come from DeepMind's own tests, mostly in lab settings.
• Demonstrations are not deployments. Folding, packing and sorting in a controlled space is not the same as working all day in a crowded hall.
• The robots shown are mostly not event robots. DeepMind's published examples use research arms, Apollo and, in a research plan, Atlas. None of the announcements we checked say that the humanoids commonly hired for events run Gemini Robotics.

What this means for event organisers

• Treat claims about AI models with care. If a supplier says a robot is powered by a foundation model, ask which model, which version and in which mode it will run on the day. Our guide to which humanoid robots you can hire today gives a realistic picture of what is available.
• On-device matters at venues. A model that runs without the cloud reduces dependence on venue networks, but DeepMind says the on-device versions are less capable than the cloud models on complex tasks.
• Expect a human in the loop. For now, most event appearances still depend on scripted or remote-controlled behaviour; see teleoperated or autonomous robots for what audiences actually see.
• Safety planning does not change. A smarter model does not remove the need for exclusion zones, trained operators and a stop procedure, as set out in our safety planning guide for public events.
• Check privacy before you book. If a robot sends camera or microphone data to a cloud model, you need to know where that data goes and what consent you need. This article is general information, not legal advice.

Frequently asked questions

What is Gemini Robotics?

Gemini Robotics is a family of AI models from Google DeepMind, first announced on 12 March 2025. It includes vision-language-action models that turn images and instructions into robot movements, embodied reasoning (ER) models that understand space and plan tasks, and on-device versions that run on the robot's own computer without a network connection. DeepMind's pages now list second-generation versions of all three.

Which robots does Gemini Robotics run on?

DeepMind has shown its models on the ALOHA 2 two-armed research platform, on Franka robot arms and on Apptronik's Apollo humanoid. Agile Robots, Agility Robotics, Boston Dynamics and Enchanted Tools were named as early trusted testers. In January 2026 Boston Dynamics and DeepMind announced joint research to bring Gemini Robotics to the Atlas humanoid, aimed first at car manufacturing.

What is Gemini Robotics On-Device?

It is a smaller vision-language-action model released on 24 June 2025 that runs locally on a robot, so it does not depend on an internet connection. DeepMind says it can be adapted to a new task with 50 to 100 demonstrations, but it performs worse than the full cloud model on unfamiliar or complex multi-step tasks. A second version is listed as being in a trusted tester phase.

Can anyone use Gemini Robotics today?

Only partly. The embodied reasoning model, now Gemini Robotics-ER 2, is available as a preview to developers through the Gemini API and Google AI Studio. The models that directly control robot movement, Gemini Robotics 2 and On-Device 2, are listed as private preview or trusted tester access with a waitlist, according to DeepMind's pages checked on 31 October 2026.

Does a robot hired for an event use Gemini Robotics?

Not necessarily, and usually there is no public evidence that it does. DeepMind's published examples use research arms, Apptronik's Apollo and, in a research partnership, Boston Dynamics Atlas. If a supplier says a robot uses a foundation model, ask which model and version it is, whether it runs on the robot or in the cloud, and who supervises the robot during the event.

How this article was made: written by the humanoid4rent editorial team with AI assistance; every factual statement was checked against the sources listed below on 2026-10-31. Spotted an error? Write to info@humanoid4rent.com.