Gemini Robotics 2 is Google DeepMind's new family of AI models for controlling physical robots, announced July 30, 2026 — and most of it isn't something you install on your phone. One piece, Gemini Robotics ER 2, is open to any developer right now in Google AI Studio; the other two need an actual robot and a spot on a waitlist.
Short answer: Gemini Robotics 2 is three models, not one app. Gemini Robotics ER 2, the reasoning layer, is live today for any developer in Google AI Studio and the Gemini API. The flagship motor-control model and Gemini Robotics On-Device 2 are restricted to robotics companies in Google's early-access program. You apply through a waitlist form, and you need physical hardware to use either one.
I spent an afternoon in AI Studio running the one piece of this family that's actually open to the public — Gemini Robotics ER 2 — and dug through DeepMind's own model cards and the Gemini API docs to confirm what the other two require. Here's what each piece does, what you can run today without owning a robot, and what it costs past the free tier.
What you'll need
For Gemini Robotics ER 2, you need nothing but a Google account. Sign in to Google AI Studio. The free tier lets you send text, images, video, and audio without a billing account attached. Want the paid tier — higher rate limits, context caching, no data used for training? You'll need a Cloud Billing account, same as any other Gemini model. For the other two models, a Google account gets you nowhere on its own. The vision-language-action model and Gemini Robotics On-Device 2 are early-access only, and DeepMind's own waitlist form asks about your organization and what robot hardware you're running before anyone gets access.
Step-by-step: how to try Gemini Robotics 2
1. Know which of the three models you actually want
"Gemini Robotics 2" gets used loosely to mean the whole family, but they're separate things. The flagship model, officially just Gemini Robotics 2, turns camera input and instructions into motor commands for a full humanoid — from walking to finger movements. Gemini Robotics ER 2 is the "brain." It plans multi-step tasks, tracks progress, and can coordinate several robots at once, but it outputs text and coordinates, not motor commands. Gemini Robotics On-Device 2 is a lighter VLA that runs locally on the robot instead of calling the cloud. Only ER 2 is something you can open in a browser today.
2. Open Gemini Robotics ER 2 in AI Studio
Go to aistudio.google.com and select the Gemini Robotics-ER 2 preview model. In the API, it's called gemini-robotics-er-2-preview for standard calls, or gemini-robotics-er-2-streaming-preview if you're using the Live API for real-time streaming.
3. Upload an image and ask it a spatial question
Drop in a photo of a cluttered desk or a workspace and ask it to point to a specific object or describe the steps a robot arm would need to pick something up. In my testing, ER 2 came back with normalized coordinates and a short reasoning trace, not just a description — that's the "embodied reasoning" part, and it's the same capability DeepMind says lets ER 2 direct real robot arms once it's wired up to one.
4. Call it from your own code
If you're building something rather than testing in the browser, the Gemini API accepts the same model ID through the standard SDK. Set thinking_level in your request — I found low settings answered faster but missed steps on multi-object scenes, while high was slower but noticeably more reliable on cluttered images.
5. Apply for early access if you have real hardware
If you run a real robot — a manipulator arm, a mobile base, a humanoid — the flagship VLA model and On-Device 2 are the ones that move it. DeepMind runs this through a formal waitlist, not open signup. Named partners already testing it include Apptronik, Boston Dynamics, Agile Robots, Franka, and Enchanted Tools, alongside 100+ other testing organizations. A form and a description of your hardware are the entry point. A credit card isn't.
6. Start narrow if you're accepted
DeepMind's own notes on On-Device 2 say it was evaluated on standing, two-armed manipulation. Not mobile bases, not full-body movement — that's what the larger flagship model is for. On unfamiliar hardware, expect it to need a few hours of demonstration data before it adapts well, per DeepMind's model card.
Example prompts you can copy
These are the kinds of prompts that got useful, specific answers out of Gemini Robotics ER 2 in AI Studio — all image or video in, structured reasoning out:
- "Point to the mug on this desk and return its normalized coordinates."
- "Look at this photo of a workbench and list, in order, the steps a robot arm would need to clear it safely."
- "Given this short video of a spill, identify what should be picked up first and explain why."
- "Two robots are in this room — a wheeled cart and an arm. Describe how they could split this task between them."
- "Identify every object in this image a robot gripper could safely pick up, and flag anything it should avoid."
The pattern that worked best: naming the physical action you want reasoned about — pointing, sequencing, dividing a task — instead of asking a generic "what's in this image" question a plain vision model would answer just as well.
Common mistakes to avoid
The biggest one I saw people make in early discussion threads: expecting ER 2 alone to move a robot. It's the planning layer. It hands off actual motor execution to a VLA model like the flagship or On-Device 2, so testing it in isolation only shows you the reasoning, not motion. Second, assuming the free AI Studio tier gives you everything — context caching and the higher rate limits are paid-tier only, and preview pricing can still change. Third, treating the trusted-tester waitlist like a normal product signup; it's built for organizations with hardware in hand, not solo hobbyists testing an idea. Fourth, skipping thinking_level in API calls and getting shallow answers on cluttered scenes. And fifth, assuming On-Device 2 handles whole-body or mobile-robot tasks. DeepMind's own limitations note says that's outside its current scope — the full model is built for that instead.
The three models compared
| Model | What it does | Who can access it | Runs where |
|---|---|---|---|
| Gemini Robotics 2 (VLA) | Full humanoid motor control, feet to fingertips | Early-access partners (waitlist) | Cloud |
| Gemini Robotics ER 2 | Plans tasks, tracks progress, coordinates multiple robots | Any developer | Google AI Studio / Gemini API |
| Gemini Robotics On-Device 2 | Lightweight motor control for standing, two-armed manipulation | Trusted Tester program (waitlist) | On the robot itself |
What Gemini Robotics ER 2 actually costs
Gemini Robotics ER 2 is free while it's in preview, up to the free tier's rate limits. Past that, I confirmed the paid-tier numbers directly on Google’s Gemini API pricing page on July 31, 2026. Input — text, image, video, or audio — runs $2.00 per million tokens. Output is $10.00 per million tokens. Context caching isn't available on the free tier; on paid, it costs $0.20 per million tokens to write plus $1.00 per million tokens per hour to store. Google hasn't published pricing for the flagship model or On-Device 2. Neither is sold per-token to the public yet, since access runs through the partner program instead.
Tools that make this easier
If you want the broader picture of what Gemini can do outside robotics, my how to use Gemini AI guide covers the consumer app, Gems, and Deep Research, and how to use Gemini in Google Docs covers the side most people actually touch day to day. For talking to a model out loud instead of reading API docs, how to use Gemini Live is the closest thing Google ships to a hands-free assistant right now. If you're building anything that calls a model from code — robotics or not — my best AI tool for code roundup and how to use Claude AI guide cover the assistants most developers reach for alongside the Gemini API. And if you're trying to figure out where Gemini's underlying models sit against GPT and Claude generally, best AI models has the current comparison.
Where this leaves you
Gemini Robotics 2 is a real jump for robotics. Whole-body humanoid control and multi-robot coordination are new. The demos DeepMind published — tying a garbage bag, screwing in a lightbulb, sorting trash — aren't cherry-picked one-shot tricks. They're the kind of multi-step chores that used to need hand-tuned scripts per robot. But if you came here without a robot in your garage, be honest about what you can actually try: Gemini Robotics ER 2 in AI Studio, and nothing more, until you're the kind of organization DeepMind's waitlist is built for.
Last updated: July 31, 2026 · By Vishal Swami, Founder & Lead AI Reviewer, AISagely
Frequently Asked Questions
Is Gemini Robotics 2 free?
Gemini Robotics ER 2 is free to use in Google AI Studio and the Gemini API up to the free tier's rate limits. Past that, the paid tier is $2.00 per million input tokens and $10.00 per million output tokens, per Google's pricing page. The full flagship and On-Device 2 models aren't sold per-token at all right now — they're limited to early-access partners.
How long does it take to get access to Gemini Robotics 2?
Trying Gemini Robotics ER 2 takes minutes. Sign in to AI Studio and pick the model. Getting access to the full flagship or On-Device 2 models depends on DeepMind's early-access review of your application and hardware. There's no fixed timeline, and access isn't guaranteed.
What's the easiest way to try Gemini Robotics 2?
Open Google AI Studio, select Gemini Robotics-ER 2, and upload a photo of a real workspace with a specific question — like asking it to point to an object or sequence a task. That's the entire model family you can test without owning a robot.
Do I need a robot to use Gemini Robotics 2?
To use the reasoning model, Gemini Robotics ER 2, no. It runs on text, images, video, and audio in a browser. To use the two motor-control models — the flagship and On-Device 2 — yes. They output commands meant to drive real robot hardware, and DeepMind's early-access program expects you to have some.
What's the difference between Gemini Robotics 2 and Gemini Robotics ER 2?
Gemini Robotics 2 is the vision-language-action model that outputs motor commands and physically controls a robot's whole body. Gemini Robotics ER 2 is the embodied reasoning model that plans the task, tracks progress, and coordinates robots — it's the layer above motor control, not a replacement for it.