Atlas just lifted a 100-pound mini-fridge using its entire body, twisted its torso 180 degrees, and set the load down without dropping a single calculation. The brain behind that move wasn't Boston Dynamics this time. It was Google. Gemini, the same AI that stunned the world as a ChatGPT rival, is now living inside humanoid robots, and the shift is bigger than anyone predicted. Atlas runs on a split-brain setup, where one side handles movement through reinforcement learning inside advanced simulations, and the other side handles thinking through two Gemini models working together.
Gemini Robotics-ER 1.6 acts as the planner, breaking a task like sorting engine parts into clear steps. Then Gemini Robotics 1.5, a Vision-Language-Action model, turns those steps into real motion. In simple terms, the robot writes its own to-do list in plain English, then moves. If a part rolls off the conveyor belt, Atlas adjusts and keeps going. Then Google did something unexpected. Gemini also powers Apollo, the humanoid from Apptronik built for warehouses and logistics floors. Apollo studies a pile of laundry, identifies each item, picks the right grasp point, and drops the clothes into the correct bin. Move the bin mid-task and Apollo notices, adjusts, and continues. It can grab soft items, oddly shaped items, even transparent objects without being pre-programmed for each one. The recycling demo went further. Apollo checks a city's recycling rules through Gemini, then sorts trash into compost, recycling, or landfill on the fly. Here is the part that should rewrite the entire robotics industry. When one robot learns a useful skill, every other humanoid running Gemini can inherit it. Artificial intelligence is no longer just thinking inside a chatbot window. AI is now walking, lifting, and learning across an entire robot fleet. And China is racing to do the exact same thing. So who actually wins the humanoid AI war?
Google shocked the world with Gemini, but now Gemini is inside humanoid robots.
Boston Dynamics is using Gemini with Atlas, a powerful humanoid that can plan tasks,
adjust during work, and recover when something unexpected happens on the floor.
Then Aptronic is using Gemini inside Apollo, a workplace humanoid built to sort laundry,
grasp tricky objects, and even understand recycling rules before touching anything.
It is helping these robots break down goals into clear steps,
turn those steps into real movement, and act without humans writing new code for every new
task. And it is now being shared across an entire fleet, so when one robot learns a useful skill,
every other robot running Gemini can become smarter too.
When Gemini launched, the whole world focused on one question,
can Google compete with ChatGPT? But that may have been the wrong question.
Google is not just trying to win the chatbot race,
it is trying to put AI into the body of real humanoid robots faster than everyone else.
And Atlas and Apollo may be the clearest sign of that strategy.
So let's start with the robots. Gemini is now powering Atlas,
the humanoid robot from Boston Dynamics. Atlas is the same robot we have seen doing backflips,
parkour, and warehouse stunts for years, but it was always running on hand-coded motion.
Now the body stays with Boston Dynamics, but the thinking part is being handled by
Google DeepMind through Gemini. Atlas works with a split-brain setup,
where one part focuses on movement and the other part focuses on decision-making.
For movement, Atlas trains inside advanced simulations using reinforcement learning.
Boston Dynamics changes things like weight, friction, and floor surface during training,
so the robot can learn how to stay stable in different real-world situations.
That training is why Atlas can squat low, twist its torso 180 degrees, and carry a 100-pound,
45-kilogram, mini-fridge using its full body instead of only depending on its hands.
Atlas understands how each joint is moving, so it can lift heavy objects with better balance,
control, and body awareness. The intelligence side is where
Gemini starts to change what Atlas can actually do. Instead of only following a fixed motion
script, Atlas uses two Gemini models to understand the goal and turn it into action.
Gemini Robotics-ER 1.6 works like the planner, so when Atlas is given a task such as sorting
engine parts, it breaks that job into clear steps. Then Gemini Robotics 1.5,
the vision-language-action model, turns those steps into real movement.
Before Atlas acts, it creates a step-by-step plan in plain English and then sends the right
commands to the body. That is why Atlas can adjust during real sequencing work,
continue when the task changes, and recover when something unexpected happens,
like a part rolling off the conveyor belt. Gemini is also now powering Apollo,
the humanoid robot built by Aptronic. Apollo is a newer humanoid designed for
real workplaces like warehouses, factories, and logistics floors,
and its body looks nothing like Atlas. But the same Gemini brain that guides Atlas
can also guide Apollo, which shows that Gemini is not locked to one robot design.
Apollo also shows why this kind of AI matters for messy real-world tasks.
It can pick up soft items, oddly-shaped items, and even transparent items without
needing every object to be pre-programmed. If someone says,
pick up the green block, Apollo can find it and reach for it. If the block moves while
the robot is reaching, Apollo keeps tracking it and changes its motion.
Most humanoid robots still need humans to write code for every new task,
but Gemini-powered robots can look at a scene, understand the goal, create a plan,
use tools like Google Search when needed, and then act.
New objects do not always need new datasets, and new tasks do not always need new code.
If one Atlas or one Apollo learns a useful skill, that skill can be shared with other
robots running the same Gemini system. This is no longer basic automation.
This is the beginning of a learning robot fleet, where one robot improves,
and the entire fleet can become smarter.
But Google is not the only one putting frontier AI inside humanoid bodies.
On the other side of the world, China is doing the exact same thing with DeepSeek.
China shocked the world with DeepSeek, but now DeepSeek is inside humanoid robots.
UBTech is using DeepSeek with Walker S2, a factory humanoid that can plan tasks,
coordinate with other robots, and work across production lines.
Then Sherry is using DeepSeek inside Mornine M1,
a customer-facing humanoid built to talk, guide, and sell products in showrooms.
It is helping Engine.ai's PM-01 and SE-01 move into public service and patrol roles.
And it is now being linked to China's Drone Swarm programs,
showing how DeepSeek is moving beyond chatbots and into real military strategy.
When DeepSeek launched, the whole world focused on one question.
Can China compete with ChatGPT? But that may have been the wrong question.
China is not just trying to win the chatbot race.
It is trying to put AI into the physical world faster than everyone else.
And humanoid robots may be the clearest sign of that strategy.
So let's start with the robots.
Let's begin with the clearest example, UBTech Walker S2.
And the reason Walker S2 matters here is not the hardware first.
It is the AI brain inside.
Walker S2 is not just walking around because someone pre-programmed every single move.
With DeepSeek R1 connected to UBTech's BrainNet 2.0,
the robot can understand tasks, plan what to do next,
and coordinate with other robots on the same production line.
So the real story starts with the AI model turning a factory humanoid
into something closer to a usable worker.
That AI brain only matters because the hardware can actually act on it.
Walker S2 is not a humanoid that only appears in polished demo videos.
Its 52 degrees of freedom allow the robot to bend,
reach, turn, grip, and move more like a worker than a basic machine on legs.
That means Walker S2 can handle parts, carry tools,
press buttons, open storage areas, move between workstations,
and take on repetitive production line tasks
A 33-pound, 15-kilogram payload means Walker S2 can carry useful factory items
and a 3-minute battery swap means the robot can keep working through long shifts
instead of stopping for a human to recharge it.
That matters because Walker S2 is already being placed where real industrial work happens.
Walker S2 is on factory floors at BYD,
Geely, and FAW Volkswagen with active trials at Airbus and Texas Instruments,
UB Tech says it has produced 1,000 units and delivered more than 500
while revenue jumped 53% year over year.
So when we talk about deep-seek inside robotics,
Walker S2 is the cleanest example.
This is not a lab moment anymore.
It is an AI model helping a humanoid understand factory work,
coordinate tasks, and keep production moving in the real world.
If humanoid robots like Walker S2 can already work almost non-stop in factories,
how long before human workers start getting replaced?
Comment below.
After Walker S2 shows deep-seek entering factory work,
Sherry's Mornine M1 shows the next step, deep-seek moving directly in front of customers.
Sherry signed a direct partnership with deep-seek in early 2025,
and deep-seek V2, V3, and R1 now power Mornine M1's voice system,
emotion recognition, and product pitch generation.
So when a customer walks into a showroom, the robot can talk,
answer questions, read basic emotional signals,
and adjust the pitch instead of repeating the same fixed script.
This is what changes Mornine M1 from a moving display into a real sales assistant.
That AI brain is wrapped in hardware built for customer-facing work,
not heavy factory labor.
Mornine M1's 40 degrees of freedom allow the robot to turn,
gesture, move its arms, guide customers, and interact more naturally
instead of just standing like a screen on wheels.
The dual-arm payload is only 3.3 pounds, 1.5 kilograms,
but that is enough for light showroom tasks like holding small items,
presenting products, and helping customers move through a sales experience.
And here is the part that makes the whole thing feel real.
Sherry has put Mornine M1 on JD.com,
China's version of Amazon, for around $41,400.
That is close to the price of a mid-tier sedan,
which makes the whole thing feel less like a secret research project
and more like an expensive product someone can actually buy.
Sherry has already deployed Mornine M1 in showrooms around the world,
including Malaysia,
but the next robot takes DeepSeek into something much more physical and public.
DeepSeek has already proven itself on factory floors and inside car showrooms,
but the next test is something harder,
a humanoid working in unpredictable public spaces.
That is exactly where Engine AI's PM-01 and SE-01 come in,
and the reason these robots matter is the AI brain behind them.
DeepSeek powers the digital employee layer behind these robots,
helping them handle interaction, reasoning,
and basic decision-making in public spaces
instead of simply following fixed movements.
That AI layer is why PM-01 and SE-01 are already being deployed
alongside police officers across Shenzhen's Nanshan and Futian districts
for patrols, public service work, and community tasks.
The AI is what makes a humanoid useful on a real street,
not just the hardware moving the limbs.
That AI brain is sitting inside hardware that became famous online for one specific reason.
Remember those viral videos of a humanoid robot
landing a clean front flip on hard concrete?
That robot was Engine AI's PM-01,
and the reason people paid attention was not just the stunt itself.
The bigger point was the level of balance,
coordination, and recovery the robot showed while moving.
PM-01 has 24 degrees of freedom,
which gives the robot the flexibility to shift weight,
stabilize itself,
absorb impact,
and recover after losing balance instead of collapsing like older humanoid robots.
That is why videos of PM-01 getting shoved,
falling,
and standing right back up spread so quickly online.
Engine AI later built a larger version called the SE-01,
and SE-01's walking style looks much closer to natural human movement
than the stiff robotic gait most people are used to seeing.
And here is the part that makes PM-01 feel real beyond the viral clips.
PM-01 starts at around $12,000,
which means universities, research labs,
and independent developers can actually afford to experiment with the robot.
So the same robot that became famous online for flipping and recovering from hard falls
is now being used in real public environments
where stability, movement, and human interaction actually matter.
Factories, showrooms, and street patrols are one thing,
but the next step is where DeepSeek enters territory
DeepSeek is now the AI system helping a combat support vehicle identify targets,
plan movement, and coordinate with nearby autonomous systems
instead of waiting for constant human commands.
The same DeepSeek models being used in factories and customer-facing robots
are now being connected to military platforms as well.
That shift is what changes the entire story.
The hardware behind that AI is Norinco's P-60 Autonomous Combat Support Vehicle.
Norinco is China's largest state-owned defense contractor,
and it unveiled the P-60 in February 2025.
Unlike the showroom robots and factory humanoids,
the P-60 was built for military coordination and battlefield support.
The P-60 can move at speeds up to 31 miles per hour,
50 kilometers per hour,
navigate terrain on its own,
and operate alongside drone swarms and robot dogs in the field.
What scares you more, humanoid robots on city patrols
or AI-powered military vehicles like the P-60?
Share your thoughts in the comments.
And here is what makes this moment so important.
Chinese officials openly described the P-60
as proof that China is rapidly closing the AI gap with the United States.
Most technical details are still classified,
but procurement records throughout 2025 show growing military adoption.
DeepSeek is no longer just a chatbot or factory tool.
It is becoming part of China's broader autonomous military ecosystem.
The P-60 is only one piece of the military picture
because the same AI brain is also being tested across the sky,
coordinating drone swarms in real time.
Instead of controlling just one drone,
the AI helps groups of drones identify threats,
divide tasks,
and keep working together in real time.
Because DeepSeek can run on board,
the swarm can continue operating even during heavy signal jamming.
During recent PLA exercises,
one operator reportedly controlled more than 200 drones at once.
Xi'an Technological University also built a DeepSeek-powered system
capable of generating 10,000 combat scenarios in only 48 seconds,
dramatically speeding up military planning and large-scale autonomous coordination.
This is where the story takes its strangest turn
because the same AI brain connected to combat vehicles and drone swarms
is also quietly moving into living rooms.
The Dream S-50 was one of the first consumer products
connected with DeepSeek R1 integration,
and the goal is to make the vacuum behave less like a timer-based appliance
and more like a system that understands the environment around it.
Instead of giving rigid commands,
users can say things like wax the master bedroom
but avoid the toys on the floor,
and the system can interpret the request more naturally.
That is what makes this part of the story feel so strange.
The same family of AI models being connected to military systems
and factory robots
is also being used to clean living rooms
and move around couches inside normal apartments.
That AI brain is wrapped in hardware
that is far more capable than older robot vacuums.
Dream's newer cleaning systems are becoming much more autonomous
than older units that simply bounced around rooms randomly.
These newer systems can map rooms,
recognize objects,
avoid obstacles,
clean edges more carefully,
and understand more natural commands
instead of relying only on simple preset routines.
And the expansion goes beyond vacuums.
Companies like Haier, Hisense, and TCL
have also started integrating DeepSeek into smart TVs,
refrigerators,
and broader home ecosystems,
pushing the AI from factories and battlefields
directly into consumer electronics.
DeepSeek is spreading from industrial robotics
into ordinary homes,
one device at a time.
What if the same AI model used for military simulations
was also inside your smart TV, fridge, or robot vacuum?
Share your thoughts below.
So why does this matter?
Because China is not just building AI models anymore,
it is building the factories,
robots,
drones,
appliances,
and supply chains needed to deploy those models at massive scale.
The United States still leads in many areas of AI research,
but China is moving faster when it comes to putting AI
into physical systems people can actually use.
DeepSeek being open source,
cheaper to run,
and connected to domestic hardware like Huawei chips
creates a feedback loop
where every robot, drone, and smart device
helps improve the same growing AI ecosystem.
DeepSeek is no longer just competing in the AI race on a computer screen.
It is moving into factories, public streets, military systems,
and even normal homes,
which is what makes this moment feel so different
from every AI wave before it.
The bigger question now is not whether these robots are coming,
because they already are.
The real question is how fast this ecosystem scales from here.
And honestly,
some of the next robots already being tested in China
make everything we just talked about look small.