Credit: Shutterstock
Google’s new AI gives humanoid robots full-body control for the first time.
A single AI model can now command a humanoid robot to walk across a room, bend down to grab an object, and place it exactly where it belongs, all without switching between separate programs.
Google DeepMind released Gemini Robotics 2, a new family of AI models built to control humanoid robots. The headline feature is whole-body control. For the first time, one system can manage a robot’s legs, torso, arms, and hands together, instead of relying on discrete programs for walking and for grasping objects.
Why Is It a Big Deal
Older robot AI systems, including Google’s own Gemini Robotics 1.5 from September 2025, split the work into two jobs. One program handled movement and balance. A second program handled picking up and manipulating objects. The two had to hand off control to each other at the right moment, which limited robots to fairly simple, tabletop tasks.
Gemini Robotics 2 removes that handoff.
In a demonstration shared by Google DeepMind, an Apptronik humanoid robot called Apollo 2 was given a simple instruction to put the watering can into the green bin on the bottom shelf. The robot walked to the table, picked up the can, walked to the shelving area, and placed the can precisely where it was told. No separate commands were needed for each step of the task.
What the Model Can Actually Control
The system also manages dexterous hand movement, not just arms. On Apollo 2, Gemini Robotics 2 controlled a five-fingered hand with 22 separate points of movement, known as a SharpaWave hand. That level of detail allows a robot to handle small or unusually shaped objects instead of using a simple gripping motion.
Google DeepMind says the model also supports two additional capabilities:
- Long-horizon task planning: The system can plan and carry out tasks that unfold across many steps in sequence, rather than a single action.
- Multi-robot collaboration: Two robots can divide a single task between them and adjust their actions based on what the other robot is doing.
Three Models, Not One
The release includes three separate but connected models.
Gemini Robotics 2 is the main vision-language-action model. It takes in camera images along with spoken or written instructions and converts them directly into motor commands.
Gemini Robotics ER 2 is the reasoning layer. It figures out the steps needed to complete a task and can explain its plan in simple language.
Gemini Robotics On-Device 2 runs directly on a robot’s onboard computer rather than a remote server, and Google DeepMind says it can be adapted to a new type of robot with only a few hours of training.
Access is limited for now. Gemini Robotics ER 2 is available through Google Cloud, Google AI Studio, and the Gemini API. The other two models, including the main system that controls full-body movement, are only available to a small group of early-access hardware partners.
The Results Are Still Uneven
Google DeepMind published performance data alongside the announcement, and it shows a technology that works but is not yet reliable. When Apollo 2 was asked to pick up objects placed on a table, it succeeded 68.4% of the time. Picking objects up from the floor worked 45.7 % of the time. Picking objects from a shelf worked 76.3% of the time.
Google DeepMind acknowledged that the robot’s movements still need to become faster and more consistent. These numbers matter because they show the real gap between a polished demonstration video and a robot performing reliably outside a controlled setting. A robot that succeeds on a task about half the time is not yet ready for daily use in a home or a busy workplace.
A New Safety Test Comes With It
Alongside the model release, Google DeepMind introduced the ASIMOV-Agentic Benchmark, a new evaluation designed to test how embodied AI systems behave in situations where their actions could cause physical harm.
As robots gain more independence to move around and handle objects on their own, testing for that kind of risk becomes a more central part of the development process, not an afterthought.
What It Means for the Field
Most humanoid robots today are built for narrow, predefined jobs. Gemini Robotics 2 is an attempt to move toward something closer to a general-purpose robot brain, one that can take a plain-language instruction and figure out how to carry it out using its entire body.
The success rates published this week show the idea works in principle, but the execution still needs work. What Gemini Robotics 2 signals is fewer robots built around narrow, hard-coded routines, and more robots guided by general models that reason through a physical task the same way an AI model reasons through a written question.
