Google's new robot AI is built to work offline. That's the barrier that's kept general-purpose robots off job sites.
Google DeepMind's Gemini Robotics 2 controls a humanoid's whole body, runs without a cloud connection, and can retrain itself for a new robot or task in a few hours. Those three specifics attack the two reasons construction has stayed a single-purpose-robot business.
Google DeepMind released Gemini Robotics 2 on July 30 — an AI model family that gives a humanoid robot whole-body control, runs on the robot itself with no cloud connection, and can be retrained for a new robot or task in a few hours instead of months [1][2]. None of that is aimed at construction. But two of those specifics — no-connectivity operation and fast retraining — are the exact reasons general-purpose robots haven't shown up on job sites while they've spread through warehouses.
What did Google actually release?
Three models, launched together [1][3]. Gemini Robotics 2 is the vision-language-action model that converts a camera feed and a spoken instruction into motor commands across an entire humanoid body — walking, crouching, reaching, and manipulating objects with a five-fingered hand, not just arm movements from a fixed base [3][4]. Gemini Robotics ER 2 is the planning layer that breaks a task into steps and coordinates more than one robot at once. Gemini Robotics On-Device 2 is the compact version built to run locally on the robot's own hardware [1].
Named partners running it include Apptronik's Apollo 2 humanoid, Franka's Duo arm, Agile Robots, and Boston Dynamics' Atlas [3]. In DeepMind's own demo, Apollo 2 picked up a watering can, walked it across a room, and set it on a shelf — a multi-step, whole-body task earlier robot models couldn't chain together [3][4].
Why does "works without internet" matter for a construction site?
Because it's the condition most warehouse robots don't have to deal with and most jobsite robots do. DeepMind says On-Device 2's offline operation matters for "warehouses and factory floors where connectivity can't always be guaranteed" [4] — that's a mild version of the problem on an active construction site, where basements, parking structures, steel and concrete decking, and remote sites routinely kill wifi and cell signal outright. A robot that needs a live cloud connection to function doesn't work in those spaces. One that runs its full decision loop onboard does.
Why does fast retraining change the economics?
Because construction's existing robots are the opposite of general-purpose, and building each one takes years. Dusty Robotics spent from 2018 to now engineering FieldPrinter, a robot that does one job — printing full-scale BIM layouts onto a subfloor at 1/16-inch accuracy, covering roughly 10,000 to 15,000 square feet a day [6][7]. Canvas spent a comparable stretch building the 1200CX, a robot that does one other job — spraying and sanding drywall joint compound, cutting finishing time from five days to two [8]. Neither robot can be pointed at a different task. That's not a limitation of those companies; it's been the only viable approach, because training a robot from scratch for each new job has been slow and expensive.
| Approach | Example | Retraining for a new task |
|---|---|---|
| Purpose-built robot (today's model) | Dusty Robotics FieldPrinter (layout) | Not possible — different task needs a different robot |
| Purpose-built robot (today's model) | Canvas 1200CX (drywall finishing) | Not possible — different task needs a different robot |
| General-purpose model (what Google just shipped) | Gemini Robotics On-Device 2 | Hours, with fewer than 200 example demonstrations [4] |
If that retraining speed holds up outside a lab, it points at a different economic model: one robot body reprogrammed across layout, material handling, or inspection tasks, instead of a separate purpose-built machine bought or rented for each one.
Is any of this coming to a job site soon?
Not yet, and Google isn't claiming it is. The VLA and On-Device models are gated behind a Trusted Tester early-access program, not generally available [4]. Every named deployment — Apptronik's Robot Park, and partner sites at Mercedes-Benz and GXO — is a factory, warehouse, or controlled test facility, not a construction site [5]. No construction robotics vendor has announced adopting Gemini Robotics 2, and a lab demo of a robot placing a watering can on a shelf is a long way from a robot navigating an active, changing, debris-covered floor.
The takeaway for a GC evaluating field robotics
Don't expect a Gemini-powered robot on your site this year. Do treat this as the signal to watch before your next equipment or robotics-vendor conversation: if a general-purpose, offline-capable robot that retrains in hours proves out beyond the demo, the pitch from vendors like Dusty Robotics, Canvas, or their competitors eventually shifts from "buy our layout robot" to "buy one robot platform, license the tasks." Ask any robotics vendor you're evaluating this year what their retraining story looks like for a second task — that answer tells you whether they're building toward this shift or building the last single-purpose machine before it arrives.
Forward this to whoever on your team owns the equipment and robotics budget. Subscribe at constructionaibrief.com.
- What is Gemini Robotics 2?
- Gemini Robotics 2 is a robot-control AI model family Google DeepMind released on July 30, 2026. It lets one AI policy control a humanoid robot's whole body — legs, torso, arms, and multi-fingered hands — instead of just the upper body, and lets multiple robots coordinate on a shared task.
- Can Gemini Robotics 2 run without an internet connection?
- Yes. One of its three models, Gemini Robotics On-Device 2, runs entirely on the robot's onboard hardware with no cloud connection required. Google DeepMind says this matters for facilities like warehouses and factory floors where network access isn't guaranteed — a description that fits most active construction sites.
- How fast can this AI adapt to a new robot body or task?
- Using a technique DeepMind calls motion transfer, Gemini Robotics On-Device 2 can adapt to a new robot body or task with fewer than 200 example demonstrations and a few hours of retraining, even when the hardware has a different shape, sensors, or joint layout than what it originally trained on.
- Is Gemini Robotics 2 being used on construction sites yet?
- No. Access to the core models requires Google's Trusted Tester early-access program, and the named deployment partners — Apptronik, Franka, Agile Robots, and Boston Dynamics — are running it in warehouse, automotive, and lab settings, not on construction jobsites.
- How is this different from existing construction robots like Dusty Robotics or Canvas?
- Today's construction robots are single-purpose machines built for one job — Dusty Robotics' FieldPrinter does floor layout, Canvas's 1200CX finishes drywall — and can't be repointed at a different task. Gemini Robotics 2 is a general-purpose control model designed to be retrained across different robot bodies and tasks, though no construction robot maker has adopted it yet.