---
title: "Google DeepMind software now controls humanoid robots that walk"
description: "Google DeepMind's Gemini Robotics 2 controls walking humanoids, five-fingered hands and robot teams. What it can actually do, and what it still can't."
dek: "Gemini Robotics 2 is Google DeepMind's software for controlling physical robots. The new version can walk a humanoid across a room, tie a knot, and hand work off to a second machine — slowly, and so far only for a handful of partners."
published_at: "2026-08-01T03:59:17.460Z"
updated_at: "2026-08-01T18:45:31.194Z"
tags:
  - "AI"
  - "Labour"
  - "Business"
source_url: "https://deepmind.google/blog/gemini-robotics-2-brings-whole-body-intelligence-to-robots/"
source_domain: "deepmind.google"
canonical: "https://hex37.com/google-deepmind-software-controls-humanoid-robots-that-walk"
---

In a demonstration video released by Google DeepMind, a humanoid robot is told to put a watering can into the green bin on the bottom shelf. It walks over to a table, picks up the can, takes a few steps to the shelving unit, lowers itself, and sets the can down where it was asked to.

Judged as a feat of physical skill, this is unremarkable. A small child does it faster and without being asked twice. What is different is that nobody wrote a program describing those steps, and nobody was standing off camera with a controller. The robot was given a sentence in English, looked at the room, and worked out the rest.

That is the claim behind [Gemini Robotics 2](https://deepmind.google/blog/gemini-robotics-2-brings-whole-body-intelligence-to-robots/), announced by Google DeepMind as the software layer meant to sit inside robots built by other companies. It is not a robot. It is the part that decides what the robot does.

## Three models, and a division of labour

The release is actually three pieces of software, and the split between them explains a lot about how this kind of system works.

The first is a vision-language-action model, or VLA — a piece of [machine learning](https://en.wikipedia.org/wiki/Machine_learning) software that takes in camera images and a written instruction and outputs motor commands: which joint moves, how far, how fast. This is the layer that produces movement. Previous versions could only drive a robot's upper body, which limited it to things you can do while standing still at a table. This one controls whole humanoids, legs included, which is why the watering can demonstration involves walking at all.

The second is called Gemini Robotics ER 2, where ER stands for embodied reasoning. It does not move anything. It looks at a scene, talks to the person giving instructions, breaks a vague request into an ordered sequence of steps, and then supervises the action model as it executes them — checking whether each step worked and re-planning if one failed. Google says it can now hold a task together across several minutes and hundreds of individual decisions, and can recognise when a task has begun and when it is finished. It is also the layer that lets two different robots coordinate on the same job, passing work between them.

The third is a stripped-down version of the action model designed to run on the robot's own hardware rather than in a data centre. That matters anywhere with no reliable internet — a warehouse basement, a field, a factory floor — and it removes the delay of sending camera footage to a server and waiting for instructions to come back.

## The hands are the hard part

Legs get the attention, but the fiddly work is at the other end. Google says the model can drive a five-fingered robotic hand with 22 independently controlled joints, and use it for things like tying a knot or sealing a ziplock bag. It can also work the plain two-pincer grippers that most industrial robots actually have, packing objects tightly into a container.

This is the sort of task where robotics has historically failed in ways that are hard to explain to anyone who has not tried it. Sealing a bag means judging pressure by sight and touch, adjusting continuously, and knowing when it has worked. Decades of research have produced machines that can do this if the bag is always in the same place and always the same bag. Doing it with an unfamiliar bag in an unfamiliar spot is the open problem.

## Learning a new body in an afternoon

Perhaps the most commercially significant claim is the least visually impressive one. Robots come in wildly different shapes, with different numbers of joints, different sensors and different reach. Traditionally, control software written for one machine is worthless on another, which means every manufacturer rebuilds the same capabilities from scratch.

Google says the on-device model can be adapted to a new two-armed robot in a few hours, typically using fewer than 200 recorded examples of the task, and it names three unrelated hardware platforms it has done this on. If that transfers reliably outside the lab, it changes the economics of the industry: hardware makers would compete on bodies while renting the intelligence, much as phone manufacturers compete on handsets running someone else's operating system.

## What the demonstrations do not show

It is worth being precise about what has and has not been shown here. This is a company blog post accompanied by curated video. The robots move visibly slowly, which Google acknowledges directly, saying movement speed still needs to advance. There is no independent testing, and the tasks shown are ones the team chose.

The unglamorous obstacles are also unresolved by better software. Robotic actuators — the motors and joints that produce force — remain heavy, power-hungry and expensive, and have improved far more slowly than the [artificial intelligence](https://en.wikipedia.org/wiki/Artificial_intelligence) directing them. A model that reasons beautifully about a doorknob still needs a wrist that can turn it. Running a large model to generate motion also introduces lag that traditional control engineering, which has spent seventy years making [robotics](https://en.wikipedia.org/wiki/Robotics) fast and predictable, does not have.

There is a broader question behind the demonstrations, and it is not really a technical one. The tasks being shown — tidying a cluttered room, packing boxes, fetching and placing objects — are the tasks that large numbers of people are currently paid to do. A general-purpose machine that learns a job from a few hundred examples, rather than from a specialist programmer, is the version of automation that has always been assumed to be decades away. It is still slow and still confined to laboratories and partner sites. But the direction of the work is not ambiguous.

Google has published a safety framework alongside the release, including a new benchmark it calls ASIMOV-Agentic. It tests whether the reasoning model will refuse an unsafe instruction from the action model, admit when a task is impossible, and ask a human for help when uncertain. It also tests whether the system notices a person approaching and stops. That last item is not a philosophical concern; it is the requirement any machine must meet before it is legally allowed to work near people.

The reasoning model is available now through Google's AI Studio developer tools and in limited preview for business customers. The models that actually move things are restricted to early-access partners. In other words: the part that thinks is public, and the part that touches the world is not — which is, at this stage, the correct order.