Google DeepMind dropped one of the most significant robotics announcements of 2026 on Wednesday: Gemini Robotics 2, a three-model AI suite that gives humanoid robots something they’ve never had before — genuine whole-body intelligence. The demo videos alone are remarkable: robots autonomously cleaning up trash, screwing in lightbulbs, tying garbage bags, and inserting cassette tapes into boomboxes. But what’s underneath the hood matters far more than the highlight reel.
From Arms to Full Bodies
The original Gemini Robotics system, released earlier this year, was impressive within limits. It could control robotic arms and hands with enough dexterity to fold origami and seal zip-lock bags. But it had a ceiling: it couldn’t direct a robot’s entire body. When a humanoid machine needs to crouch to pick something off the floor, step around an obstacle, or maintain balance while manipulating an object overhead, upper-body control alone falls apart.
Gemini Robotics 2 tears down that ceiling. The flagship model in the suite takes over the whole machine — legs, torso, arms, and hands — optimizing the robot’s center of gravity in real time to minimize the risk of falls. In benchmark testing, the system achieved a 92% success rate at unscrewing a lightbulb, a task that requires precisely coordinated full-body movement and fine-grained hand control at the same time. That number isn’t just impressive on paper; it represents a qualitative leap in robot reliability for real-world deployment.
A Three-Layer Architecture
Google DeepMind built Gemini Robotics 2 as a layered system designed to mimic how humans think and act. At the top sits Gemini Robotics ER 2 — ER standing for Embodied Reasoning. This is the model that does the thinking.
ER 2 takes a natural language description of a task — “clean up the kitchen counter” — and breaks it down into a high-level plan with potentially hundreds of individual steps. It watches continuous video from the robot’s cameras to track progress, can identify the last successfully completed step if something goes wrong, and knows when a task is actually finished. That last point sounds trivial but isn’t: a robot that can’t reliably tell when it’s done will either stop too early or keep working past the point of completion, wasting time and potentially creating problems.
The model also supports what Google calls “tool calling,” meaning ER 2 can reach out to external services — like Google Search — to clarify ambiguous instructions before acting on them. If a prompt is unclear, the robot asks rather than guesses.
Below ER 2 sit two Vision Language Action (VLA) models that translate high-level plans into actual motor commands. The primary one is Gemini Robotics 2 itself, handling whole-body humanoid control. The second is Gemini Robotics On-Device 2, a lighter model designed to run directly on a robot’s onboard computer. DeepMind says this on-device variant can be adapted to an entirely new robot body within a few hours of training — a critical feature for manufacturers who want to deploy the technology across different hardware platforms without months of retraining.
Multi-Robot Collaboration
One of the more quietly remarkable capabilities of ER 2 is its ability to split tasks across multiple robots. A complex chore that might take one robot several minutes can be distributed among a team, with ER 2 coordinating the effort like a project manager assigning subtasks. According to Google, the system can manage chores that span hundreds of steps and take multiple minutes to complete — far beyond the reach of previous robotic AI systems.
This matters enormously for industrial and logistics applications, where the value of robotic workers scales with how many tasks they can tackle in parallel, not just how well they handle a single operation.
Honest About Limitations
Google was notably candid in its announcement about what the demo does and doesn’t show. While the company described the robots as acting “fully autonomously,” it also acknowledged that each showcased task was specifically trained in advance using human teleoperation, video examples, and simulations. The robots aren’t general-purpose machines that can walk into any environment and figure things out on the fly.
That transparency stands in pointed contrast to some competitor announcements. Elon Musk has repeatedly described Tesla’s Optimus humanoid robot as potentially “the biggest product ever,” projecting $10 trillion in eventual sales. But a 2025 target of having 1,000 Optimus robots working in manufacturing facilities was missed, and projections of 50,000 to 100,000 units deployed in 2026 are also off-track.
Google isn’t claiming Gemini Robotics 2 is ready for unstructured real-world environments. What it is claiming — and demonstrating — is a reproducible, measurable improvement in humanoid robot capability at a level of autonomy no one had publicly achieved before.
Safety at the Hardware Level
Physical AI systems carry stakes that software-only AI doesn’t. A language model that generates a wrong answer is annoying; a robot that misunderstands a command can cause physical injury or property damage. Google is treating this seriously.
Carolina Parada, head of robotics at Google DeepMind, told Wired: “The safety question is even more pressing because you’re putting them in a lot of other situations. There’s a lot of uncertainty that will show up, and so you want to be able to understand the safety question more deeply.”
Each layer of the Gemini Robotics 2 architecture includes its own guardrails. Beyond that, Google is simultaneously releasing ASIMOV-Agentic, a new benchmark specifically designed to evaluate whether a humanoid robot’s AI can recognize when a command could lead to a harmful outcome — before it acts. Named with an obvious nod to Isaac Asimov’s Three Laws of Robotics, the benchmark is an attempt to make robot safety evaluation as rigorous and reproducible as model capability benchmarks.
Physical AGI: A New Goalpost
Google engineers Steven Hansen and Peng Xu wrote in the accompanying blog post: “By watching continuous video feeds, robots can now track their own progress, adapt if something goes wrong, and know exactly when to move on to the next step.”
Google DeepMind has said openly that the long-term goal is “physical AGI” — a robot capable of performing any task a human can. That’s an audacious claim, and the company isn’t suggesting Gemini Robotics 2 gets it there. But it does move the goalposts. The previous question was whether AI-powered robots could do useful things at all. The new question is how quickly the gap between trained demonstrations and true generalization can be closed.
Availability and What Comes Next
Gemini Robotics ER 2 is available now for developers via Google Cloud, the Gemini API, and Google AI Studio. Google is rolling out access to early hardware partners, with the full model suite — including the whole-body VLA and on-device variant — available through partner programs.
Consumer-facing robots remain a long way off. But for enterprise robotics teams working in manufacturing, logistics, healthcare support, and research, Gemini Robotics 2 represents the most capable foundation yet for building machines that can actually work.
The race toward robots that can do meaningful, multi-step physical work is accelerating faster than most people anticipated. Google just made it significantly clearer who’s leading.
Sources: Google DeepMind Blog, SiliconANGLE, TechBriefly, Bloomberg
