Gemini Robotics 2: Google DeepMind Lets Robots Work as a Team
On this page
Google DeepMind released Gemini Robotics 2 on July 30, 2026 — a three-model system that lets a single humanoid control its whole body (legs, torso, arms, and fingers) from one policy, and lets two separate robots split up a chore and finish it together. I’ve been tracking the robotics-model space since the original Gemini Robotics launch last year, and this is the first release from Google that reads less like a lab demo and more like an actual product roadmap: an on-device model for factories with spotty Wi-Fi, a cloud reasoning model any developer can call today, and a motor-control model that already runs on real hardware from two different manufacturers.
What Gemini Robotics 2 actually is
“Gemini Robotics 2” isn’t one model — it’s three, and each does a different job in the stack:
- Gemini Robotics 2 (VLA) — the vision-language-action model that turns a camera feed and a spoken instruction into motor commands. The headline detail: one checkpoint drives Apptronik’s Apollo 2 humanoid with two different hand designs (SharpaWave and Inspire) and also the Franka Duo bi-arm research platform with a Robotiq gripper, without retraining per robot body.
- Gemini Robotics-ER 2 — the embodied-reasoning model, essentially the “brain” that talks to a human, reads a cluttered room, and plans a multi-step task that might take several minutes to finish.
- Gemini Robotics On-Device 2 — a compressed VLA built to run locally on the robot itself, so a warehouse arm or an inspection robot keeps working even if the network connection drops.
The party trick DeepMind is showing off is whole-body control: a robot that can crouch to pick something off the floor, stretch to reach a high shelf, and walk across a room, all under one continuous policy instead of separate “walking” and “grasping” modules stitched together. Layered on top is dexterity fine enough to screw in a light bulb or tie a knot, and — the part I find genuinely new — multi-robot collaboration, where two robots divide a single job (say, clearing a table) instead of each one working a job in isolation.
Why “whole-body” and “multi-robot” are the hard parts
Most humanoid demos over the past few years have quietly cheated: the legs handle locomotion, a separate arm-control stack handles manipulation, and a human operator (or a scripted state machine) hands off between the two. That’s fine for a trade-show video, but it breaks the moment a robot needs to crouch while reaching, or catch its balance while lifting something heavy. A single policy that owns the whole body removes that seam, which is exactly why DeepMind is leading with it.
Multi-robot collaboration is the harder problem underneath the surface, though. Two robots dividing labor means each one has to model what the other is doing, avoid stepping on each other’s task, and adjust in real time if its partner falls behind — closer to a coordination problem than a control problem. If Google can make that reliable outside a demo, it’s a bigger deal for warehouses and factory floors than any single-robot trick, since most real work (loading a truck, stocking a shelf) is naturally a two- or three-worker job.
Who can actually use it today
This is the part that separates a real release from a research paper: access is tiered, and only one piece is open right now.
- Gemini Robotics-ER 2 is in public preview — any developer can call it through the Gemini API or Google AI Studio today.
- Gemini Robotics 2 (VLA) and On-Device 2 are gated behind DeepMind’s Trusted Tester Program, which means an application form and, presumably, actual robot hardware to run it on.
That’s a sensible split. The reasoning model is pure software — safe to hand to any developer with an API key. The motor-control models drive physical machines that can move fast and hit people, so a controlled rollout to vetted partners (Apptronik among them) makes sense before it opens further.
Quick reference
| Model | Role | Availability (as of July 30, 2026) |
|---|---|---|
| Gemini Robotics 2 (VLA) | Whole-body motor control, dexterity | Trusted Tester Program only |
| Gemini Robotics-ER 2 | Embodied reasoning, multi-step planning | Public preview via Gemini API / AI Studio |
| Gemini Robotics On-Device 2 | Local, offline-capable motor control | Trusted Tester Program only |
| Confirmed hardware | Runs the same VLA checkpoint | Apptronik Apollo 2 (two hand variants), Franka Duo bi-arm |
Where this fits next to Google’s other July announcements
Robotics wasn’t the only thing Google shipped this week. The company also confirmed Gemini 3.6 Flash and the start of Gemini 4 pretraining, and it’s now under a binding EU order to open Android to competing assistants. Seeing all three land in the same stretch tells me Google is running its “one model family, every surface” strategy pretty literally: the same Gemini reasoning core is showing up in phones, in the cloud API, and now in robot bodies, just wrapped in different action layers. That’s a meaningfully different bet than a company that ships a robotics model as a one-off side project.
It’s also worth reading against the wider AI-agent trend I’ve been covering — tool-calling standards like the recent MCP spec update are about software agents calling software tools. Gemini Robotics 2 is effectively the same idea — a model reasoning about a task, then choosing actions — except the “tool” is a robot arm and the API call is a torque command. The industry is converging on the same architecture (a reasoning layer plus an action layer) whether the output is a database query or a screwed-in light bulb.
The obvious caveat
Everything above the “public preview” line is a controlled demo, not a shipping consumer product. DeepMind has shown whole-body control and two-robot handoffs work in its own testing; it hasn’t shown either running unsupervised at scale, on a factory floor, for a full shift. Robotics companies have a long history of demo videos that look further along than the underlying reliability actually is (Boston Dynamics knows this territory well). I’d treat the reasoning model as genuinely available now, and the motor-control claims as “promising, gated, and worth revisiting once Trusted Testers start publishing their own results.”
FAQ
What is Gemini Robotics 2?
It’s Google DeepMind’s July 2026 robotics release: a set of three AI models — a whole-body vision-language-action model, an embodied-reasoning model, and an on-device model — that together let humanoid and bi-arm robots plan, move, and manipulate objects, and coordinate with other robots on shared tasks.
Can I use Gemini Robotics 2 right now?
You can use the reasoning half, Gemini Robotics-ER 2, today through the Gemini API or Google AI Studio. The motor-control models (the VLA and On-Device versions) are restricted to DeepMind’s Trusted Tester Program and require applying for access plus compatible robot hardware.
What robots does Gemini Robotics 2 run on?
DeepMind has confirmed the same VLA checkpoint driving Apptronik’s Apollo 2 humanoid (in two different hand configurations) and the Franka Duo bi-arm platform — meaning one trained model, not a from-scratch version per robot.
How is this different from the first Gemini Robotics model?
The original Gemini Robotics, released last year, focused mainly on tabletop manipulation with arms. Gemini Robotics 2 extends control to the entire body (legs and torso included), adds finer dexterity for delicate tasks, and introduces multi-robot collaboration, none of which were part of the first release.
Sofia follows emerging technology, from AI and VR to IoT and blockchain, and translates the hype into plain language. She cares about what these tools mean for everyday users, not just the headlines.
