Gemini Robotics 2: How Google's AI Gives Humanoids Full-Body Control
Google's Gemini Robotics 2 lets humanoid robots reason, balance, use dexterous hands, and coordinate with other robots on shared tasks.
What is Gemini Robotics 2?
Gemini Robotics 2 is Google’s latest AI model for controlling humanoid robots, built to handle whole-body movement, fine hand dexterity, and coordination between multiple robots working on the same task. Rather than programming a robot for one specific motion, Google is training a single generalist model that reasons about a goal, plans the physical steps, and controls the robot’s full body, from feet to fingertips, to carry it out. The company demonstrated the system on two robot platforms, referred to in the demos as Apollo and Duo, performing tasks like packing bags, unscrewing a light bulb, tying off a trash bag, and tidying a garage together.
TL;DR
- Whole-body control is one of the three pillars of the release, letting a humanoid coordinate balance, legs, torso, and arms as a single system instead of executing isolated pre-scripted motions.
- Dexterous manipulation covers tasks Google says are notoriously hard for robots, including sealing a Ziploc bag, unscrewing a light bulb from a socket, and tying a knot in a trash bag using 22 separate joints in a robotic hand.
- Multi-robot collaboration allows two robots, each running its own copy of the same model, to divide a task, hand off responsibility mid-task, and communicate about when help is needed.
- The system pairs an embodied reasoning model that interprets scenes and language with a vision language action model that turns decisions into precise motor commands.
- Google frames this as solving the “missing piece” of robotics: AI models that can generalize across tasks rather than hard-coded routines built for one repeated motion.
- Demonstrations included recovery behavior, where a robot recognizes a failed grasp or unstable stance and retries rather than stalling out.
- The tasks were deliberately chosen to be hard for machines but trivial for humans, exposing how much hidden coordination goes into everyday actions like packing a bag or picking an object off the floor.
How does Gemini Robotics 2 actually work?
Google describes the system as two connected models working together. The first is an embodied reasoning model that takes in what the robot sees through its cameras and interprets a natural language instruction, something like “pack the kids’ sports bags” or “tidy the garage and put the tools back in their kit.” This model figures out the sequence of subgoals: what needs to happen first, what depends on what, and when a task is actually finished.
The second piece is a vision language action model, which takes those subgoals and converts them into the low-level motor commands a robot body needs to execute them. That means driving dozens of joints and actuators at once, from the legs and torso down to individual fingers, while keeping the robot balanced. Google’s team notes that if a robot leans too far forward, the model has to command the legs to compensate in a fraction of a second, the kind of adjustment humans make without thinking but that requires deliberate computation in a machine.
The split matters because reasoning and motor control are different problems. Reasoning is about sequencing and judgment: what order to do things in, what counts as “done,” when to ask another robot for help. Motor control is about translating that judgment into physical torque and position changes across a body with many degrees of freedom. Gemini Robotics 2 is built to handle both inside one pipeline.
Why is whole-body control such a hard problem?
Humans move in a coordinated way without thinking about it. Reaching for an object while standing involves shifting weight, adjusting balance, and moving an arm all at once, and the body handles that automatically. Robots don’t have that automatic coordination built in. Google’s engineers point out that a humanoid has to make many separate decisions across its entire body just to complete what looks like a simple reach-and-grab motion.
This is different from the specialized humanoid robots people have seen running, jumping, or doing backflips in viral videos. Those systems are typically tuned for a narrow set of physical feats. Gemini Robotics 2 is aimed at generality instead: one model that can pack a bag, screw in a bulb, tie a knot, and hand off a task to another robot, without being rebuilt for each new job. Google’s framing is that a generalist model adds more real-world value than a specialist one, because a single robot that can do many different things is more useful in a home or workplace than one that does a single trick very well.
What can the robots’ hands actually do?
Manipulation is where the demos get specific. Google highlights tasks that go beyond simple pick-and-place: sealing a Ziploc bag around packed food, unscrewing a spherical light bulb from its socket, and tying off a trash bag with a knot. Each requires precise fingertip contact, an understanding of 3D space, and multi-step motion (twist, push, pull) rather than a single grasp-and-release action.
Built like a system. Not vibe-coded.
Remy manages the project — every layer architected, not stitched together at the last second.
The light bulb task is called out specifically because the object is round, meaning the contact points for the fingers have to be calculated carefully to keep grip. The trash bag knot task was reportedly considered close to impossible by some team members during development, since a robotic hand can have around 22 separate joints to coordinate, far more than the simple open-close motion of a basic gripper. Google also notes it is pushing dexterity even for robots equipped with simpler parallel grippers, which still need precision and spatial understanding to reorient objects and place them accurately, since the “last centimeter” of a placement can matter for whether a task actually counts as complete.
How do multiple robots collaborate on one task?
The third pillar is multi-robot coordination. In one demo, two robots are told to tidy a garage: grab a scrubbing mitt and spray bottle together, carry them to a bin, then kit up all the tools and put the kit away. Rather than a single neural network controlling both robots, each robot runs its own independent copy of the same model and does its own reasoning. They communicate about task state, meaning one robot can recognize its portion is finished and hand off the remaining work to the other.
In the garage example, the first robot completes its part of the task and then calls on the second to finish kitting and closing the tools. Google’s team describes this as orchestration through reasoning: each robot decides for itself when to act and when to step back, rather than following a fixed, pre-scripted division of labor. The goal, according to Google, is that robots capable of collaborating with each other can take on a much wider range of tasks than any single robot working alone.
Is this actually useful outside a demo?
The tasks shown, packing a lunch, unscrewing a bulb, tying a trash bag, tidying a garage, are mundane by design. That’s the point: they’re trivial for a person and genuinely hard for a machine, which makes them a reasonable proxy for whether a robot can generalize. The demos also included visible failure and retry behavior, where a robot recognizes an unsuccessful grasp or unstable step and adjusts rather than getting stuck, a signal that the system is meant to work in unpredictable, real-world conditions rather than a controlled lab setup.
Whether this scales into robots doing useful, unsupervised work in homes or workplaces still depends on speed, reliability, and cost, none of which were detailed in Google’s demonstrations. What Gemini Robotics 2 does show is a shift in emphasis: from robots as pre-programmed machines that repeat one action, to robots as reasoning systems that plan, adapt, and coordinate with each other.
Frequently Asked Questions
What is Gemini Robotics 2 used for?
It’s an AI model designed to control humanoid robots, handling whole-body balance and movement, precise hand manipulation, and coordination between multiple robots working on a shared task.
How is Gemini Robotics 2 different from other robot demos like backflipping humanoids?
Those systems are typically built for a specific physical feat. Gemini Robotics 2 is designed to be generalist, meaning one model can handle many different everyday tasks (packing bags, unscrewing bulbs, tidying rooms) instead of one specialized motion.
Can multiple robots really work together with this system?
Yes. In Google’s demos, two robots each ran their own copy of the same model and coordinated through reasoning, dividing a task, handing off remaining work, and communicating about task completion rather than being controlled by one shared network.
What makes tasks like tying a knot or unscrewing a bulb hard for robots?
- ✕a coding agent
- ✕no-code
- ✕vibe coding
- ✕a faster Cursor
The one that tells the coding agents what to build.
These tasks require fine motor coordination across many joints (a robotic hand can have around 22), precise fingertip contact, and 3D spatial understanding, all of which are far more complex than a basic gripper’s open-close motion.
Does Gemini Robotics 2 work on any robot body?
Google demonstrated it on humanoid platforms with both dexterous multi-fingered hands and simpler parallel grippers, showing the model adapting its control strategy to the hardware available.