Jul, 134 min to read
Frontend
News · Google DeepMind splits its robotics push into a full-stack model and a programmable reasoning model
Mar, 124 min to read
FrontendGoogle DeepMind splits its robotics push into a full-stack model and a programmable reasoning model
Gemini Robotics and Gemini Robotics-ER take two different postures toward the developer, one closed-loop and one open to your own code.
Two models, two different surfaces for the builder
Google DeepMind announced two models built on Gemini 2.0. The first, Gemini Robotics, is described as its most advanced vision-language-action model, meant to let robots perform a wider range of real-world tasks. The second, Gemini Robotics-ER, is positioned differently: it lets roboticists run their own programs using Gemini's embodied reasoning.
That distinction is the most concrete thing in the announcement. One model is the whole loop — perception, language, and action bundled together. The other exposes reasoning as a component you build around. Google is not shipping a single robot brain; it is shipping two entry points with different amounts of control handed to the developer.
Vision-language-action as a single interface
The framing of Gemini Robotics as a vision-language-action model matters for how a robot gets instructed. It folds seeing, understanding a request, and producing motion into one system. For a builder, that collapses what used to be separate stages — a perception pipeline, a planner, a controller — behind one call.
The tradeoff is legibility. When the model owns the full path from image and instruction to action, the seams where you would normally inspect or override are inside the model. The announcement gives no detail on how much of that internal behavior is observable, which is exactly the question a team integrating it would need answered.
Gemini Robotics-ER and the case for your own control loop
Gemini Robotics-ER is the version that respects existing engineering. Roboticists keep their own programs and call the embodied reasoning as a service inside them. That is the frontend that fits teams who already have controllers, safety layers, and hardware abstractions they are not going to throw away.
This split acknowledges something real about robotics work: the last-inch control code is specific, tested, and often safety-critical. Offering a reasoning model that plugs into it, rather than replacing it, is a more honest fit for how robots actually ship than a monolithic action model alone.
Trusted-tester status is the current constraint
Both models are being tested with partners and trusted testers. There is no open access, no pricing, and no stated hardware compatibility in the announcement. Anything a team plans around these models today is provisional until that access widens.
The specific implication is this: the meaningful decision Google has surfaced is not which model is more capable, but which posture you want — an end-to-end action model or a reasoning layer inside your own code. For anyone building on top, Gemini Robotics-ER is the one that leaves your control loop intact, and that is the choice worth watching as access opens.
Find this story relevant to you?
Contact us to find a unique solution
Need an AI engineering partner that can actually build?
We help businesses integrate AI, build AI-powered products, automate high-value workflows, and modernize the software systems behind them.
Related reading
More analysis around product delivery, operational AI, and the systems work that makes deployment hold up in reality.