Google DeepMind released Gemini Robotics 2 on July 30, giving humanoid robots the ability to coordinate full-body movements—walking, crouching, and manipulating objects—in one continuous action. The suite includes an on-device model that can run locally on robotic hardware, a detail that matters for Windows-powered industrial environments where cloud dependency is a nonstarter.

The release, demonstrated on Apptronik’s Apollo 2 humanoid, marks a shift from tabletop arm movements to what DeepMind calls “feet to fingertip” control. A vision-language-action model translates camera input and natural language into motion, letting a robot walk to a shelf, pick up a watering can, and place it in a bin without requiring step-by-step instructions.

But the most significant piece for enterprise IT is Gemini Robotics On-Device 2. It’s a lighter model that runs right on the robot, eliminating the latency and connectivity risks of cloud AI. Google says it can be adapted to a new dual-arm robot design with fewer than 200 real-world examples and only a few hours of training. That lowers the barrier for manufacturers who run Windows-based control systems on the factory floor and need AI to operate reliably offline.

What actually changed: three models for whole-body, planning, and local execution

Gemini Robotics 2 isn’t a single model but a stack designed for different parts of a robot’s decision-making load:

  • Gemini Robotics 2 is the primary vision-language-action (VLA) model. It handles full-body motion for humanoids and dual-arm systems, converting camera inputs and spoken instructions into coordinated movements.
  • Gemini Robotics ER 2 adds high-level embodied reasoning. It plans multi-step tasks that can last minutes, monitors whether each step succeeded, and corrects failures on the fly. It also enables multi-robot collaboration—different robot types can share information and split work.
  • Gemini Robotics On-Device 2 runs locally on robot hardware. It’s optimized for situations where cloud access is unreliable or latency matters. Adaptation to new robots requires minimal data and time.

In one demo, Apollo 2 walks to retrieve a watering can and stows it on a lower shelf. The robot crouches, reaches, adjusts its center of gravity, and uses its hands—all parts of a single coordinated task. That’s a leap from earlier Gemini Robotics releases, which centered on upper-body manipulation.

DeepMind published performance figures that show both progress and rough edges. With Apollo 2 equipped with Inspire hands, the model succeeded at lifting objects from a table 68.4% of the time, from the floor 45.7%, and from a shelf 76.3%. Fine dexterity is uneven: unscrewing a lightbulb reached 92% success, but screwing one in hit just 36%. Those numbers matter for anyone planning real deployments—the system is not yet a plug-and-play generalist.

What it means for you: edge-ready AI and safety benchmarks

For Windows users, the immediate impact depends on where you sit.

If you’re an IT administrator in manufacturing or logistics, Gemini Robotics On-Device 2 starts to make humanoid robots plausible for brownfield environments. Many factory floors run Windows IoT or Windows Server on industrial PCs that control conveyor belts, robotic arms, and automated guided vehicles. Local AI that can be tuned to new hardware with only a few hours of data collection means you could add a humanoid to a line without rewiring the whole network for cloud AI. The model’s cross-platform design—it was demonstrated on both Apollo 2 configurations and Franka Duo dual-arm systems—suggests Google is trying to avoid hardware lock-in.

If you’re a safety officer or plant manager, DeepMind’s release of the ASIMOV-Agentic safety benchmark is a signal that safety is being baked into the evaluation of these models. The benchmark measures a robot’s ability to refuse unsafe actions, recognize when a task can’t be completed safely, and request human help. Gemini Robotics ER 2 also includes human-proximity detection that triggers a safe stop when someone enters the workspace. But these are software guards, not physical barriers. They don’t replace risk assessments, light curtains, zone scanners, or lockout procedures. A robot with whole-body motion can make whole-body mistakes, and software-only safety hasn’t been proven reliable enough to skip hardware safeguards.

If you’re a developer or system integrator, Gemini Robotics ER 2 is available now through Google AI Studio and in limited preview on the Gemini Enterprise Agent Platform. That gives you a way to test high-level task planning and human-robot communication without needing a physical robot. The full Gemini Robotics 2 and On-Device 2 models are restricted to early-access partners, so broad experimentation is still ahead.

If you’re a home user, this release is mostly a curiosity—no consumer robot uses Gemini Robotics 2 today, and Windows PCs won’t suddenly run humanoid control stacks. But as AI moves to the edge, the techniques proven here may eventually trickle down to Windows-based robotics toolkits or simulators.

How we got here: from tabletop arms to walking humanoids

Google’s robotics work has steadily expanded its scope. Earlier Gemini Robotics models focused on upper-body manipulation—arms picking and placing objects on a tabletop, for example. That approach required the robot to stay stationary and work within a constrained workspace.

The collaboration with Apptronik, announced earlier this year, set the stage for whole-body control. In July, Apptronik opened Robot Park, a facility where Apollo 2 robots collect training data for next-generation AI models. That data pipeline feeds directly into Gemini Robotics 2. The Apollo 2 runs an Nvidia GPU stack and is designed for industrial tasks, making it a plausible target for Windows-based control loops.

DeepMind’s move also reflects a wider industry push toward general-purpose humanoids. Tesla’s Optimus, Figure’s 01, and Boston Dynamics’ Atlas all aim for whole-body manipulation. But Google’s approach to making the reasoning and motion layers separable—then offering an on-device variant—addresses a real barrier: the cloud-dependent models used by many competitors can’t guarantee the sub-100ms response times needed for safe human-robot interaction.

What to do now: steps for Windows administrators and safety teams

If you’re responsible for a Windows-based edge environment that could one day include humanoids, here’s what to do today:

  1. Evaluate the on-device model’s requirements. Google says adjustment to new dual-arm robots needs fewer than 200 training examples and a few hours of work. That suggests you could spin up a prototype on existing hardware without a massive data-collection campaign. Check whether your industrial PCs can run the model’s inference stack—Google hasn’t published precise hardware requirements, but similar edge models often require a GPU with at least 8 GB of VRAM.

  2. Test high-level planning in Google AI Studio. Even without hardware, you can experiment with Gemini Robotics ER 2 to see how it reasons about multi-step tasks. This can help you gauge whether the model’s planning logic aligns with your workflows before investing in a physical pilot.

  3. Apply for early access if you have a concrete use case. Google is offering Gemini Robotics 2 and On-Device 2 to selected partners. If your facility has a repetitive task that could benefit from a mobile manipulator—restocking, kitting, or machine tending—an application could put you ahead of the curve.

  4. Don’t treat ASIMOV-Agentic as a safety certification. The benchmark is a tool for researchers, not an industry standard. Consult your regional occupational safety authority’s guidelines for collaborative robots and ensure any deployment includes physical safeguards like force-limited joints, speed limits, and monitored safe stop zones. Windows-based safety configurators can integrate real-time risk assessment, but the AI model’s own safety checks are one layer among many.

  5. Watch for real-world validation. The performance numbers on dexterity—92% for unscrewing a lightbulb versus 36% for screwing it in—show the model still struggles with consistent fine manipulation. Wait for independent testing before depending on it for tasks that require high first-time success rates.

Outlook: a foundation, not a finished product

Gemini Robotics 2 is a research milestone, not a commercial turnkey system. The on-device model’s adaptability and the safety benchmark signal that Google is thinking about real-world constraints. But the wide gap between lab demos and factory floors means the next six months will be about pilot programs, not broad rollouts.

For Windows-focused enterprises, the key question is when open APIs or local SDKs let you integrate Gemini Robotics 2 directly into existing Windows-based robot controllers. Google’s track record with TensorFlow and MediaPipe suggests that on-device deployment pathways will expand—but they aren’t there yet. In the meantime, the release resets expectations for what humanoid robots can do with AI whole-body control, and it puts pressure on competitors to match the on-device capability that Windows edge deployments require.