Gemini Robotics 2 brings AI agents into real motion

What this is about

Google DeepMind introduced Gemini Robotics 2 on July 30, 2026. The new model family is meant to let robots not only understand instructions, but control movement across the whole body: walking, crouching, grasping, carrying objects, and collaborating with other robots.

That matters because many robotics demos still depend on narrow, prepared tasks. DeepMind is trying to shrink the gap between language models, world understanding, and motor control.

What Gemini Robotics 2 actually does

DeepMind describes three building blocks. Gemini Robotics 2 is a vision-language-action model that turns visual and language input into movement. Gemini Robotics ER 2 handles planning, communication, and progress tracking as an embodied reasoning model. Gemini Robotics On-Device 2 is designed to run locally on robots and adapt to new robot bodies with a few hours of data.

The examples include Apollo 2 humanoids and dual-arm platforms. Tasks include picking up a watering can and placing it on a lower shelf, unscrewing a light bulb, tying a bag, and packing items into kits. Gemini Robotics ER 2 is available in Google AI Studio; the action models remain limited to early-access partners.

Why it matters

For consumers, the obvious question is whether the household robot is finally getting closer. For companies, the question is when flexible automation becomes more useful than fixed programmed cells. Both questions depend not only on better models, but also on safety, speed, cost, hardware, and liability.

DeepMind gives concrete success rates that keep the story grounded. Some pick-and-place and tool tasks score highly, including 89.6 percent on precise insertion tasks using a Franka Duo platform. Multi-finger tasks are more mixed: unscrewing a bulb reaches 92 percent, screwing one in reaches 36 percent, tying a bag reaches 44 percent, and closing a ziplock bag reaches 40 percent.

That is why the announcement is interesting: it shows progress, but also why robotics is harder than chat. A mistake is not just a wrong answer. It can be a broken object, a blocked workflow, or a safety risk.

In plain language

Imagine someone who does not just memorize a recipe, but cooks in an unfamiliar kitchen. They need to find drawers, hold the pot, use the stove safely, and notice when another person is in the way. Language understanding alone is not enough.

Gemini Robotics 2 is a step in that direction: the AI is meant not just to say what should happen, but to plan and execute movement in space.

A practical example

A small warehouse processes 2,000 spare-part boxes per day. Today, a human handles exceptions: parts lying sideways, open boxes, missing labels. A robot with a fixed routine can handle only standard boxes. A system like Gemini Robotics 2 could theoretically see that a box is half open, ask a second robot to hold it, grip the part, close the lid, and move the box onward.

If the success rate for one subtask is only 40 percent, however, deployment is not automatic. The workflow then needs a fallback station, human approval, or a narrower set of reliable tasks.

Scope and limits

First, the best demos do not equal production readiness. DeepMind itself says movement speed and human-level dexterity still need improvement.

Second, hardware remains an open question. A model that works on Apollo 2 or Franka Duo will not necessarily work reliably on every low-cost industrial robot or household device.

Third, safety is the real bottleneck. DeepMind introduces ASIMOV-Agentic as a new benchmark and emphasizes stops when humans come close. Still, every real environment needs physical safeguards, risk analysis, and clear responsibility.

SEO & GEO keywords

Google DeepMind, Gemini Robotics 2, Gemini Robotics ER 2, humanoid robots, Apptronik Apollo 2, Vision-Language-Action, robotics safety, ASIMOV-Agentic, On-Device AI, Physical AI