Gemini Robotics is Google DeepMind’s family of AI models for controlling robots. As of October 19, 2026, you can use only one of them yourself: Gemini Robotics ER 2, the reasoning model, which is in preview on the Gemini API and in Google AI Studio. The two models that actually move a robot’s motors, Gemini Robotics 2 and Gemini Robotics On-Device 2, are available only to early-access partners, according to DeepMind’s July 30, 2026 announcement.
That split matters if you are deciding whether to build on it. ER 2 looks at images, video and audio and returns text: object coordinates, plans and function calls. Your own robot software, or a separate action model, still has to turn that output into motion.
The current family has three models
DeepMind released the second generation on July 30, 2026. Each model does a different job:
- Gemini Robotics ER 2 (embodied reasoning): plans multi-step tasks, points to objects, tracks task progress in video and calls tools or robot APIs. Its model card says it is built on Gemini 3.5 Flash. Open to developers.
- Gemini Robotics 2: a vision-language-action (VLA) model that turns camera input and an instruction into motor commands for humanoids and two-armed robots. Early-access partners only.
- Gemini Robotics On-Device 2: a smaller VLA that runs on the robot without a network connection and can be adapted to a new robot body. Early-access partners only.
In DeepMind’s design, ER 2 acts as the planner and hands steps to an action model. It can call a Gemini Robotics VLA, another company’s VLA, or ordinary robot functions you write yourself. The same idea in other labs’ work is covered in our guide to Physical Intelligence and openpi, where π models handle the action side.
What ER 2 can do through the API
The Gemini Robotics ER documentation lists five capability groups:
- Spatial reasoning: points to objects, draws bounding boxes, tracks objects in video and plans trajectories. Points come back as [y, x] coordinates normalized to 0–1000.
- Agentic vision: uses code execution to crop, zoom or annotate an image before answering, which helps with small objects and gauge reading.
- Task orchestration: breaks an instruction such as “put the apple in the bowl” into locate, grasp and place steps and calls your robot functions in order.
- Streaming: a separate endpoint takes continuous audio and video through the Live API for low-latency robot agents.
- Video progress: finds the moment an event happens in a video and judges whether a task step succeeded.
There are two model codes. Use gemini-robotics-er-2-preview for standard requests and batch jobs, and gemini-robotics-er-2-streaming-preview for real-time Live API sessions. Both accept text, images, video and audio, take up to 131,072 input tokens and return up to 65,536 output tokens.
How to start with ER 2
- Create a Gemini API key in Google AI Studio or the Google Cloud console.
- Add restrictions to the key. The documentation says requests from unrestricted keys are refused with a 403 error.
- Try a prompt in AI Studio first: upload a photo of a table and ask the model to point to each object.
- Call the model from the Google GenAI SDK (Python, JavaScript, Java or Go) or REST with the model code above. The official overview page includes a working pointing example.
- Connect the output to your robot: pass the coordinates to your motion planner, or register your robot’s functions so ER 2 can call them during orchestration.
The model is a preview, so its behavior and model code can change. Its predecessor, Gemini Robotics ER 1.6, was shut down on August 31, 2026, according to the Gemini API deprecations page. If old code still names gemini-robotics-er-1.6-preview, switch it to an ER 2 model code.
Per-token prices for ER 2 are listed on the Gemini API pricing page. Check that page before you run long video jobs, because video input and high thinking levels raise both cost and response time.
Rules that apply when a robot uses it near people
Two conditions in the documentation affect real deployments.
- Safety is your responsibility. Google states that the model can make mistakes and that the developer must keep the area around the robot safe. A wrong coordinate sent straight to a motor can damage equipment or hurt someone, so keep your own limits, stop conditions and speed caps between the model and the hardware.
- Privacy consent comes first. The documentation’s privacy notice says you must not let identifiable people be near the robot while it sends video and audio to the model unless they have been told and have consented that Google may use that data under the Gemini API terms. Google also asks developers to reduce personal data, for example with face blurring.
Can you get the VLA models?
Not as an individual developer today. DeepMind’s announcement offers Gemini Robotics 2 and On-Device 2 only to early-access partners and does not list a download, a public API endpoint or a price. The robots shown with Gemini Robotics in DeepMind’s material include Apptronik’s Apollo humanoid and Boston Dynamics’ Spot, but appearing in a demo is not the same as a product you can buy with the model installed.
If you need an open VLA you can run on your own hardware now, look at openpi or NVIDIA’s GR00T models instead. For DeepMind’s own published results, our Korean-language analysis of Gemini Robotics 2 whole-body benchmarks separates the lab numbers from what has been shown on real robots.
How it fits with other physical AI models
Gemini Robotics sits in the “robot brain” layer. It is not a simulator and not a world model that generates training video; NVIDIA covers that layer with Cosmos. For the wider picture of how models, data and hardware connect, start with what physical AI is, and for the companies building robots that might run these models, see our list of humanoid robot companies in 2026.