Google's New AI Robot Brain Is INSANE….

Infinity Robotic
4 followers
AI Generated·May contain errors
Video Summary
AI GeneratedGoogle is revolutionizing robot training by using the Gemini 4 Argon model to translate visual inputs—such as human videos or rough sketches—into physical skills. By utilizing a \"video-to-simulation-to-real\" workflow, the AI creates physics simulations for robots to practice in before executing tasks in the real world, significantly speeding up the development of complex movements.
Google just taught a humanoid robot a new skill by showing
its AI a video of a person playing a backyard game.
The robot is Apollo, and the game is Cornhole.
But the surprising part isn't that Apollo can throw a beanbag.
Google's Gemini 4 Argon watched the human demonstration,
worked out how the robot should move,
and helped build a physics simulation to develop the throw.
Apollo then landed seven consecutive shots in a real-world test.
And Google didn't stop there.
The same model turned a rough drawing into a full-body robot movement,
controlled a robotic dog,
and helped another robot improve through thousands
of simulated practice sessions.
These experiments point to a very different way of teaching robots.
The story begins with Apollo,
a humanoid developed by the American robotics company
Aptronic. Google DeepMind has been working
with Apptronik to give robots more capable intelligence,
and in October, it showed what happens
when a powerful general-purpose Gemini model enters
the process of teaching a robot a new physical skill.
The researchers chose Cornhole.
The rules are simple.
A player stands some distance from an angled wooden board
and tosses a small bag toward a hole near the top.
For a person, the motion feels natural.
You look at the target, adjust your stance, move your arm,
and release the bag at just the right moment.
But a humanoid has to calculate how its joints, hand,
and entire body work together to make that throw.
Instead of starting with a motion programmed by engineers,
Argon received raw video of a human playing.
It analyzed the scene, estimated the positions of
the objects, and worked out the body's movement
in three-dimensional space.
Then, it translated that understanding
into movements suitable for Apollo.
A human arm and a robot arm do not move in exactly the same way,
so copying each frame isn't enough.
The model had to reason about the robot's own joints
and the physical constraints of the task.
Argon also generated code for a physics simulation.
Think of it as a practice ground where a robot's movements can
be tested against rules for gravity, motion, and contact.
The simulation helps engineers explore what happens
when a bag is released too early,
when the arm follows a different path,
or when the robot's weight shifts.
During a throw, those details matter because even a small change
at the hand can send the bag far away from its target.
The skill was then transferred to a physical Apollo robot,
and in Google's published demonstration,
the robot sank seven consecutive shots.
That result matters
because the starting point was an ordinary human video,
not a long set of handwritten commands for the exact throwing motion.
The company calls this kind of workflow, video-to-simulation-to-real.
It is a way to move from watching an action
to developing a skill that can be tested in
the physical world. And there is an important distinction here.
Argon wasn't acting as a motor
that directly powered every joint.
It was helping build the instructions, simulation,
and control behavior that made the movement possible.
That's what makes the experiment interesting
If more unfamiliar tasks can be approached this way,
developing a robot's next skill might become much faster.
But Google's next experiment removed the human video entirely.
This time, the input was a rough pen and paper drawing.
It showed an Apollo humanoid holding a ball
and throwing it towards six boxes stacked in a pyramid.
From that sketch,
Argon constructed a virtual scene with a robot, table, ball, and targets,
It then worked out a complete throwing motion for a simulated humanoid
with 76 degrees of freedom, including its body and five-fingered hands.
That movement involved much more than swinging an arm.
The robot had to grip the ball, prepare its stance,
accelerate through the throw, release at the right point.
and maintain balance afterward.
Argon generated and checked the sequence
in a MuJoCo physics environment.
The simulated ball struck the stack and knocked down all six boxes.
This particular result happened in simulation,
not on a physical Apollo.
Still, it shows how far the starting point has changed.
A simple drawing became a detailed robot movement
that could be tested against physical rules.
And Argon isn't limited to humanoids.
In another demonstration,
Google used Boston Dynamics' four-legged Spot robot,
which carries a robotic arm.
Spot was asked to pick up soda cans and put them into a
trash can with a foot pedal.
That sounds like one small chore, but it creates a problem.
The robot has to stand on three legs
while pressing the pedal with the fourth,
keep its balance, and move the can into the opening at the same time.
Argon built a simulated version of the trash can,
so the robot could practice
the pedal movement before trying it on the actual bin.
The demonstration
also used checks to verify that Spot's foot lined up with the pedal,
and its hand reached the opening before releasing the can.
In the footage, Spot presses the pedal and drops the can inside.
This is a very different body from Apollo,
yet the same model contributed to solving the physical task.
Google showed another example with a much simpler robotic arm.
A mug had fallen onto its side, and a crumpled napkin was nearby.
Argon used camera observations
and available movement commands to get the arm to stand
the mug upright and place the napkin inside.
Unlike the Apollo throwing experiment,
This demonstration focused on observing a scene,
choosing actions, and checking the result across several steps.
Together, the examples show how one reasoning model
can help across very different robotics setups.
But the most important experiment may be the one where
Argon helped improve a robot that already knew how to move.
Researchers started with a single photograph of a tabletop workspace.
The model reconstructed that scene as a physics simulation,
including the shapes and positions of objects.
It then generated more than 4,500
practice episodes across different tasks.
The resulting synthetic data was used to further train
an existing vision-language action model,
which turns visual input and instructions into robot movements.
In Google's test,
the physical robot's success rate rose from 56% to 75%.
Its average completion time also fell from 35 seconds to 24 seconds.
That is a 19 percentage point improvement in success rate,
achieved by adding experience created inside a simulation.
The results come from a particular research setup,
but they demonstrate why generated training data is so valuable.
Physical robots can be slow and expensive to train.
A simulated workspace can produce many more examples
without resetting real objects after every attempt.
Then came a smaller robot arm called SO-101.
Its job was to pick up an object and place it accurately in a box.
On the first attempt shown by Google, the task took 58.9 seconds.
Argon reviewed the result and refined the behavior.
The next attempt to
It took 22.9 seconds, although the arm grazed the rim of the box.
After another improvement, it completed the task in 17.4 seconds
and placed the object near the center.
The striking part is that the model was using the outcomes
of earlier attempts to help make the next attempt better.
This also helps explain why Argon matters when Google already
has Gemini Robotics 2 and its embodied reasoning model.
ER2. Those models are built specifically for robots.
ER2 can understand a task, break it into steps,
work with movement controllers,
and monitor whether an action succeeded.
Gemini Robotics 2 converts visual observations
and instructions into physical actions.
Argon overlaps with these abilities,
but Google has also demonstrated it writing simulation code,
reconstructing environments, creating training data,
and improving robot control behavior.
In other words, the new model can help develop
the skills that a robot will later perform.
And Google is exploring another path with Genie 4,
its interactive world model.
In a recent research demonstration,
Argon and Genie worked together to simulate Apollo playing checkers.
Argon broke the task into instructions,
while Genie generated the changing scene as the action unfolded.
The robot wasn't physically moving checkers on a real board.
The point was to test how an AI planner could guide an interactive
generated world, opening up new ways to explore robot tasks.
Before taking them into reality.
This fits into Google's wider robotics effort.
Its partnership with Apptronik includes Apollo robots collecting
real-world training data at Robot Park and customer sites.
The company is also working with Boston Dynamics
on bringing Gemini-based intelligence to robots
such as Atlas. Real-world experience,
dedicated robot control models, physics simulations,
and general purpose reasoning are becoming different parts
of the same research program.
Each addresses a different problem, how to understand a task,
learn the movement, execute it,
and improve when the outcome isn't right.
Apollo's seven cornhole shots make a great demonstration.
But the bigger story is the path
from a human video to a new physical skill.
Google is showing that a powerful AI model can help robots practice,
generate training experience,
and refine their movements instead of relying
on a separate hand-built program for every task.
There is still a long distance between
these controlled experiments and a robot that can
learn anything on demand, but the direction is clear.
What happens when teaching a robot becomes
as simple as showing it what to do?
Let me know what you think in the comments,
and subscribe to Next Gen Humanoids for more
updates.
