Google DeepMind's Gemini Robotics 2: Whole-Body Control, Multi-Robot Teams, and Fast Adaptation

Google DeepMind has released Gemini Robotics 2, a suite of three AI models designed to give humanoid robots something they’ve largely lacked until now: the ability to think, move, and work together in the real world without being explicitly programmed for every situation.

This is a meaningful step forward from the original Gemini Robotics, which launched in March 2025 and focused primarily on upper-body control. The new suite tackles the full robot, multi-robot coordination, and the problem of transferring skills to new hardware bodies quickly.

The Three Models

Gemini Robotics 2 (VLA) is the vision-language-action model at the heart of the suite. It takes in what the robot sees and hears, and converts that into physical movement. The significant upgrade here is whole-body control: where the previous model directed arms and hands, this one controls the entire humanoid, from legs and torso through to fingertips.

That means a robot can now step sideways, squat down, or bend to reach something in a cluttered space, dynamically adjusting its balance as it goes. It also supports complex 22-degree-of-freedom hands, enabling finer manipulation. Google’s demo footage shows robots cleaning up rubbish, screwing in a lightbulb, inserting a tape into a boombox, and tying a garbage bag. The company says this is real-time, fully autonomous operation.

Gemini Robotics ER 2 is the embodied reasoning model, built on Gemini 3.5 Flash with a 128k token context window and support for text, image, video, and audio. Think of it as the high-level brain that sits above the VLA. It handles multi-step planning, communicates with humans in plain language, and can call external tools like Google Search or Google Calendar to inform its decisions.

A few details worth noting: ER 2 can track task progress using the robot’s own camera feed. If a robot makes an error mid-task, ER 2 identifies the last correctly completed step and resumes from there, rather than starting over. It can also “think ahead” while the robot is still moving, so planning and action happen in parallel rather than sequentially.

Gemini Robotics On-Device 2 is the efficiency-focused variant, built on Google’s Gemma on-device models and designed to run locally on the robot itself rather than in the cloud. The headline capability here is fast adaptation: it can learn to operate a completely new robot body with fewer than 200 examples and just a few hours of data. That’s a significant reduction in the time and cost typically involved in retraining robotic AI for new hardware.

Multi-Robot Collaboration

One of the more practically interesting additions is coordinated multi-robot operation. The system allows different robots to communicate with each other, understand each other’s physical capabilities and limitations, and autonomously divide up tasks based on which robot is best suited to each part.

This matters because many real-world tasks are genuinely too complex or physically demanding for a single robot. Having robots that can negotiate a shared workload without human micromanagement opens up a broader range of deployment scenarios, particularly in logistics, warehousing, and industrial settings.

Safety Gets Its Own Benchmark

Google is also releasing ASIMOV-Agentic, a new benchmark for robotic AI safety. It tests three specific behaviours: whether the reasoning model refuses unsafe commands from the VLA, whether it can recognise when a task is physically impossible, and whether it proactively asks a human for help when uncertain.

In Google’s own lab tests, ER 2 detected a person entering the robot’s working area and directed Apollo 2 to put down the objects it was holding and move into a safe resting pose before resuming the task once the person had left. The benchmark itself is available on HuggingFace, so any organisation building robotic AI can use it to evaluate their own systems, not just DeepMind’s.

Availability

Gemini Robotics ER 2 is available now via Google AI Studio and in private preview on the Gemini Enterprise Agent Platform. The VLA and On-Device models are available to early-access partners.

What Does This Mean for You?

If you work in robotics, manufacturing, logistics, or are building products on top of robotic platforms, the ER 2 model is probably the most immediately accessible piece of this. It’s available through the Gemini API today, which means you can start experimenting with high-level task planning and tool-calling for robotic systems without needing physical hardware in front of you.

The On-Device 2 model is the one to watch if you’re thinking about deploying robots across different hardware configurations. The ability to adapt to a new robot body in hours rather than weeks changes the economics of robotic AI significantly.

For everyone else, consumer-facing applications are not on the near-term horizon. Google is clear that this is still early-stage research and deployment. But the trajectory is notable. DeepMind describes this as progress towards “physical AGI”, meaning a robot that can do anything a human can. The gap between that goal and current performance metrics (a 45.7% success rate picking objects off the floor, for instance) gives a reasonable sense of where things actually stand.

The foundations are getting more solid. The benchmarks are being published openly. And the tooling is becoming accessible to developers before the hardware reaches scale. That sequencing is worth paying attention to.