QWE AI Academy Hot take: the piece of Gemini Robotics 2 you actually get to touch this week isn’t the humanoid-walking VLA that’s blowing up your feed. It’s the boring-sounding reasoning model sitting next to it. And that’s the piece worth learning.
On July 30, 2026, Google DeepMind dropped Gemini Robotics 2 – a system that lets humanoids walk, crouch, and manipulate objects while reasoning through tasks, unlike the previous model which only controlled a robot’s upper body. The demo everyone’s sharing shows an Apptronik Apollo 2 fetching a watering can. Cool. But if you’re a developer reading this, you can’t call that model from your laptop tonight. You can call its sibling. Here’s how, and where the sharp edges are.
The problem: the shiny demo isn’t the thing you can use
Every article you’ll read today leads with the walking humanoid. Fair – it’s the marketing moment. But the release actually contains three separate models, and only one of them ships to the public API today.
The lineup, fast: Gemini Robotics 2 is the VLA – full humanoid motor control, walks and crouches, gated to partner companies. Gemini Robotics On-Device 2 is a smaller local VLA for embedded hardware, also partner-only. Gemini Robotics ER 2 is the embodied-reasoning layer that sits above both – plans multi-step tasks, coordinates multiple robots, hands off execution. That last one? Publicly available via the Gemini API and AI Studio as of launch day. On-Device 2 adapts to entirely new robot bodies in a few hours of training data, per the DeepMind announcement – but you still can’t call it from your laptop.
Guess which one is available in Google AI Studio right now? Only ER 2. The VLA – the one doing the walking – stays gated with partner robotics companies. If you don’t have an Apollo 2 in your garage, that model doesn’t exist for you yet.
Why the news-recap tutorials fall short
Most posts today will list the three models, replay the watering-can clip, and call it a day. That’s fine if you want a summary. It’s useless if you want to build something.
The actionable path – using ER 2 as a planner brain over a robot you already have or a simulated one – is buried in the docs. So let’s pull it out.
What Gemini Robotics 2 actually gives you (as a developer, today)
Here’s the honest breakdown of what you can and can’t do the day this shipped:
| Model | What it does | Can you use it today? |
|---|---|---|
| Gemini Robotics 2 (VLA) | Full-body motor control for humanoids | No – partner access only |
| Gemini Robotics ER 2 | Planning, spatial reasoning, video understanding, tool calls | Yes – Gemini API + AI Studio |
| Gemini Robotics On-Device 2 | Local VLA for embedded deployment | No – partner access only |
Think of ER 2 as the high-level brain. Pass it a camera image, ask it what’s graspable – it returns normalized 2D coordinates and labels. Ask it to plan a five-step cleaning task – it produces an ordered sequence your motor stack can consume. The VLA translates those steps into joint torques; ER 2 decides what the steps are. If you’re building with a simple gripper API already, ER 2 slots directly on top.
That division of labor is worth sitting with for a second. “Whole-body control” sounds like one capability, but it’s actually two completely separate engineering problems: knowing what to do next (perception, planning, task decomposition) and knowing how to move joints to execute it (motor control, proprioception, real-time correction). ER 2 owns the first problem entirely. The VLA owns the second. Most developers building on top of existing robots already have something handling the second – which is exactly why ER 2 is the more immediately useful piece.
How to actually try ER 2 in the next 10 minutes
You don’t need a humanoid robot. You need a browser and a Google account.
- Open aistudio.google.com. Sign in.
- Click Get API key in the left sidebar, then Create API key. AI Studio is free to use with no subscription required – production billing via the Gemini Developer API kicks in only when you move beyond prototyping, and no credit card is needed to start.
- In your code, use the model ID
gemini-robotics-er-2-preview– not the older 1.6 model, which is going away at end of August 2026. - Feed it an image from your robot’s camera (or any photo of a scene) and ask it to point to objects, plan a task, or detect whether an action succeeded.
Here’s a minimal call in Python, using the standard Gemini SDK:
from google import genai
client = genai.Client(api_key="YOUR_KEY")
response = client.models.generate_content(
model="gemini-robotics-er-2-preview",
contents=[
{"role": "user", "parts": [
{"inline_data": {"mime_type": "image/jpeg", "data": image_bytes}},
{"text": "Point to every object I could pick up. Return normalized 2D coordinates and a label for each."}
]}
]
)
print(response.text)That’s it. The model returns coordinates you can pipe straight into your own motor control code. Per the Gemini API docs, ER 2’s pointing, task planning, and video-progress capabilities are the core output types – coordinates are a first-class result format, not a workaround.
The stuff nobody’s telling you
Three sharp edges the launch posts glossed over.
1. Two endpoints exist, and picking wrong wrecks your latency. The streaming variant – gemini-robotics-er-2-streaming-preview – runs through the Live API and is built for continuous audio and video input at low latency. The standard gemini-robotics-er-2-preview is a request-response model. Feed a 30fps webcam stream to the standard endpoint and you’ll feel the difference immediately. The API docs call this out, but it’s easy to miss on first read.
2. If you built anything on ER 1.6, you have a deadline. Swap model="gemini-robotics-er-1.6-preview" to model="gemini-robotics-er-2-preview" (or the streaming variant) before end of August 2026 – that’s when 1.6 gets shut down. Silent deprecations like this bite hobby projects that haven’t been touched in a month.
3. Pricing is a fog right now. ER 2 is built on Gemini 3.5 Flash under the hood – the docs confirm this, which suggests billing will land somewhere in the Flash tier. But as of July 30, 2026, a dedicated ER 2 line item hasn’t appeared in the official pricing table. Prototype on the free tier and treat any cost estimate as a guess until Google publishes numbers.
A real-world example: the watering-can task, decomposed
The demo everyone’s watching: Apollo 2 given the prompt “put the watering can into the green bin in the bottom shelf” – walks to the table, picks up the can, crosses the room, places it precisely. One fluid motion on screen. Architecturally, it isn’t one thing.
ER 2 owns the “what to do” layer (scene → plan: navigate, grasp, navigate, place). The VLA – running on Gemini Robotics 2 proper – takes each step and converts it into the joint torques that make legs move and fingers close. Turns out the same model checkpoint ran across three different embodiments in the demo: Apollo 2 with SharpaWave hands, Apollo 2 with Inspire hands, and Franka Duo with a Robotiq gripper. Same ER 2 call; different bodies executing downstream.
The safety piece is worth calling out specifically. According to the Google blog, ER 2 halts a humanoid when a person enters its proximity and only resumes once the area is clear – not a configured rule, but an output of the Safety Instruction Following benchmark improvements baked into training. For anyone deploying near people, that’s not a footnote.
Pro tip: Send each step of a multi-step task as a separate prompt rather than one giant instruction. It keeps the model focused on the current state, and lets you inject sensor updates between steps. This is how the official examples are structured, and it’s the difference between a plan that survives contact with the real world and one that doesn’t.
Where the hype meets reality
Community reaction to the earlier Gemini Robotics launch framed it as a “ChatGPT moment for robotics” – though, as InfoQ’s roundup noted, some argued the real moment arrives when regular people can buy robots for personal use. Today’s release doesn’t get us there. It gets developers one step closer.
What it does do: it lowers the bar for anyone who wants a smart planner sitting on top of a dumb robot. Which is most robots.
The more open question – the one worth watching – is what happens when the VLA access gate opens up. Right now, ER 2 is the public piece and the VLA is partner-only. When that flips, or when On-Device 2 shows up in the API, the developer story changes substantially. Until then: ER 2 is the actual product.
FAQ
Can I run Gemini Robotics 2 on my Raspberry Pi robot?
Not the whole-body VLA – partner-only. But ER 2 is an API call, so your Pi can hit it over the internet as a planner and run your own control loop locally. That works today.
Do I need a robot at all to learn this?
No – and this is actually a useful way to start. Take a photo of any cluttered desk. Pass it to ER 2 in AI Studio and ask it to return the normalized coordinates of every graspable object. Then ask it whether a task looks “completed” based on a before/after image pair. You’re testing the exact output format you’d consume in real hardware code, without any hardware. Spatial reasoning, video understanding, multi-step planning – all of it is useful on static inputs first. Hardware comes later.
How is this different from just using Gemini 2.5 Pro with a robot API?
A common assumption – but ER 2 isn’t just a general model pointed at robots. It’s tuned on embodied-reasoning benchmarks, returns pointing coordinates as a first-class output format, and does video progress classification natively (detecting mid-execution failures like spills and slips from raw video, not snapshots). For casual scene description, 2.5 Pro is fine. For anything that actually needs to orchestrate a robot – multi-step tasks, multi-robot coordination, safety-aware halting – ER 2 is the right tool.
Next step: open AI Studio, generate a key, and run the pointing example above with a photo from your phone. Ten minutes. If it feels underwhelming compared to the humanoid videos, that’s the point – the humanoid videos aren’t the product you can use yet. The pointing API is.