Humanoid robots just had a crazy week.
We got Moya, the warm-in robot from
China that looks like it was built to
make people uncomfortable. Boston
Dynamics Atlas is now moving heavy
objects and showing signs of real
factory intelligence. Agibot's full-size
humanoid is playing autonomous table
tennis. Cody is bringing child-friendly
robots into museums, hotels, and maybe
healthcare. And Alibaba just dropped
Quen robot, an AI brain built to control
the next generation of machines. So,
let's start with the thing that got the
most reaction this week. And for good
reason. Droid up, officially known as
Shanghai Robotics, just unveiled their
full body humanoid robot called Moya.
And the demo immediately split the
internet between fascination and pure
discomfort. Here's the context. Droidup
is a Shanghai based startup that's only
about 2 years old. They raised $28.5
million in total and have already burned
through nearly half of it on payroll and
recruitment bonuses for their 32
employees. Their main venture finance
partner is itself only worth around $50
million. So the financial picture is
well slim. But despite that, they
unveiled Moya at Janguan Robotics
Valley, which is quickly becoming the
center of China's humanoid robot race.
And the demo caught a lot of attention
for obvious reasons. The robot itself,
Moya, stands about 5 1/2 ft tall, weighs
around 70 lb, and runs on a platform
called Walker 3, the same skeleton that
helped an earlier droid up robot
finished third in the world's first
humanoid half marathon. But Moya isn't
built for speed or endurance. She's
built to sit across from you and make
you forget she's not a person. Behind
each eye, there are working cameras. And
those cameras don't just track movement.
They track your face, your expressions,
and Moya mirrors them back. She smiles,
nods, narrows her eyes like she's
actually listening. Underneath the
silicone skin, Droidup added padding
designed to mimic human fat and muscle
with a structure that resembles a rib
cage. The skin itself runs warm, 90 to
97° F. That's actual human body
temperature built into the machine. So,
when you reach out and touch her hand,
you're not getting cold plastic. You're
getting something that feels
disturbingly close to a real person.
Droidup also redesigned the internal
structure around an artificial spine
instead of rigid joints, letting Moya's
torso twist and bend the way a human
spine actually distributes force. That's
not just an aesthetic update that tells
you this company is actively iterating
towards something more convincingly
alive, not just maintaining a one-time
stunt. The main claim Droid is pushing
is that Moya walks with 92% accuracy
compared to a natural human gate. That's
their own number, not independently
verified. Watching the footage, it does
look noticeably less robotic than most,
more like a deliberate, careful human
step. That said, Reddit's r/S
singularity had a field day with it. One
of the most upvoted comments called it a
natural geriatric gate, and people were
pointing out the word natural in the
marketing was doing some seriously heavy
lifting. The broader online reaction
mixed genuine skepticism with the kind
of humor that subreddit is known for.
One user broke down the financials in
detail. Less than $15 million left for
actual development, a leadership team
with zero successful businesses under
their belt, and a primary investor worth
only $50 million. The promo video
editing didn't help Droidup's
credibility either. The orange juice
pouring scene was clearly cut to hide
what the robot can and can't actually
do, and multiple people noted the hands
look like wooden sticks. The YouTube
review commentary echoed the same
criticisms. The face described as
moderately realistic at best, kind of
rigid and hard with something about the
shape of the facial features that just
doesn't land as human. That said, there
is a live event video from a few months
back where actual attendees walked up to
Moya at a physical venue, touched her,
and interacted with her directly. And
those unscripted reactions are genuinely
more interesting than any polished
promo. Droidup is pricing Moya at
$173,000
and pointing her at hospitals, elder
care facilities, banks, museums, and
train stations, places where a human
used to greet you, guide you, or just
sit with you. The demo even showed Moya
pouring a glass of orange juice for an
elderly woman, which is precisely the
elder care use case they're pitching.
They're also offering a customizable
version with configurable appearances,
and some of those configurations are
very clearly aimed at companionship
rather than customer service. First,
production batch is targeting around 50
units with a wider release planned for
later this year. Now on the completely
opposite end of the credibility
spectrum, Boston Dynamics and Atlas. A
new report from KB Securities analyst
Kang Sun Jin says Atlas is approaching
the level of autonomy actually needed
for industrial deployment. And the
reason this is being taken seriously is
what's backing it up. Boston Dynamics,
which is Hyundai Motor Group's US-based
robotics subsidiary, can now run
simulations equivalent to millions of
hours of robot training in a single day.
Not weeks, a single day. And the skills
that come out of those simulations can
be transferred to a physical Atlas robot
in roughly 1 hour. That's a development
cycle that was basically unthinkable
even 2 years ago. Boston Dynamics also
recently ran a campaign using Atlas in a
football context to specifically
showcase human robot interaction and
physical AI, which tells you something
about how confident they're getting in
public facing demonstrations. A lot of
this progress comes directly from
strategic partnerships. Google DeepMind
on the AI and learning system side and
NVIDIA providing the high performance
computing infrastructure for large-scale
simulation. Together, that combination
is compressing timelines significantly.
There's also something worth
understanding about Atlas's hardware
design that makes everything faster.
Most humanoid robots are nightmares to
simulate accurately because they have
dozens of different actuator types and
complex mechanical systems that are hard
to model virtually. Atlas cuts through
that with only two actuator types across
the entire body. Arms and legs are also
symmetrically designed, which simplifies
both the control system and the
simulation. They eliminated cables
running across joints as well, enabling
continuous joint rotation, reducing
maintenance, and giving the robot
movement flexibility that conventional
humanoids can't match. The practical
results are genuinely impressive. Atlas
moved a refrigerator weighing over 100
lb despite only being trained on loads
of 50 to 70 lb. Boston Dynamics says
tasks like that require more than visual
perception. Full body coordination,
weight anticipation, balance, and
precise force control, all working
simultaneously. The handstands and back
flips Atlas is known for aren't just
showboating either. Those athletic
movements are specifically used to train
balance, agility, slip recovery, and
endurance that directly translate to
industrial environments. KB Securities
Project's Boston Dynamics could capture
15% of the global humanoid robot market
by 2035 and as much as 60% of the
premium industrial humanoid segment.
Bold numbers, but given where Atlas
actually sits right now, not without
foundation. On a completely different
capability axis, Agibbot dropped
something that honestly didn't get the
attention it deserved. Their full-size
bipeedal humanoid robot, the Yuan Jang
A3, just completed a fully autonomous
table tennis match. No remote control,
no scripts, no human intervention at any
point. That might sound like a novelty,
but table tennis is genuinely one of the
tougher benchmarks in robotics. Balls
are moving at over 5 m/s with spin,
trajectory, and landing points shifting
in fractions of a second. The robot has
to run a complete closed loop system
continuously processing visual
perception, predicting trajectories,
planning whole body movements, executing
precise strikes, and fluidly switching
between offense and defense, all without
stopping. This came out of a
collaboration between Agibbot and a team
led by Jang Shangh.
They built on something called Spike
Ping Pong, the world's first table
tennis motion control algorithm
specifically for humanoid robots. Paired
with a 20 kHz highfrequency spike camera
developed by Professor Hang Tun, that
combination delivers visual response
speeds 10 times faster than conventional
camera solutions with millimeter level
prediction of exactly where the paddle
needs to meet the ball. Agibbot is
calling this the first time a full-size
bipeedal humanoid has completed a match
with fully autonomous decision-making
from start to finish. And the
demonstration backs that up cleanly. On
a much smaller scale, literally, there's
a project out of Seattle worth knowing
about. A company called Mind Children
built a humanoid robot named Cody. And
it's a completely different kind of
robot from everything else we've
covered. Cody is 3 feet tall with
glowing hazel eyes, expressive facial
animations, and a design specifically
built to be approachable for kids in
public facing environments. The target
venues are museums, hotels, and art
galleries where Cody would assist
visitors with information, offer
companionship, support learning
activities, and help organizations
create more engaging experiences. What
makes it more than a cute mascot is that
it operates fully autonomously. No
remote human operator in the loop.
Onboard sensors, computing, and AI
software handle navigation,
conversation, and basic independent
decision-making in real time. Near-term
development goals include enabling
simple physical tasks like pressing
buttons, and a second generation version
is already scheduled for 2027. Longer
term, mine children is targeting
healthcare and education deployments.
The AI backbone here is Hyperon, a
decentralized AGI framework from
Singularity. It's designed to go beyond
rule-based automation through knowledge
representation, reasoning engines, and
what the company describes as human
inspired motivation systems, letting the
robot prioritize tasks, pursue goals,
and adapt its behavior as circumstances
change. On the funding side, they've
been crowdfunding and have pulled in
over $600,000 toward a $1 million
target, which tells you there's a real
audience paying attention. Rounding
things out on the infrastructure side,
Alibaba just launched Quinn Robot, their
first embodied AI model family, and it's
a genuinely significant move. Developed
by Tongi Lab and already in pilot
testing with selected Alibaba Cloud
Enterprise clients, the suite addresses
a fundamental gap in robotics. A vision
language model can understand a command
like go to the kitchen, find the red
cup, pick it up, and place it on the
shelf. But understanding a task and
actually performing it are two entirely
different things. Connecting language
and visual understanding to physical
motor control is hard. Partly because
robot training data comes in completely
different formats from internet data and
is expensive to gather. Mixing data
sources carelessly creates conflicts
instead of improvements. So Alibaba
split the problem into three specialized
models. Quenroot Nav handles movement
and navigation. In a demo on a Unitry GO
two quadriped running NVIDIA Jetson Thor
hardware with just a single
lowresolution camera, it navigated an
unfamiliar apartment following spoken
instructions across multiple rooms with
no preloaded maps, maintaining an
inference latency of 196 milliseconds.
Quenroot Manup handles physical
interaction, grasping, moving, and
manipulating objects trained on over
38,000 hours of open- source data. And
it recently topped the generalist
category at the Robo Challenge realworld
robotics benchmark with a process score
of 59.83
and a task success rate of 45%. Quen
robot world acts as the world model,
predicting how environments change and
helping robots reason about the likely
outcomes of their actions before taking
them. Alibaba also released Quen Robot
Claw, an agent framework, letting the
models use the robot suite as physical
world tools. One demo had an agent
locate a restroom, spot an outof order
sign on the door, and independently
reroute to a different location with
zero human input. They also open sourced
Chat 2 Robot, a browserbased platform
for testing embodied AI interactions,
which is a solid move for the developer
community. And this is why Alibaba's
move matters. The US has Deep Mind,
Nvidia, Figure, Skilled, and physical
intelligence pushing robot brains. But
China has the hardware army, Unitry,
Agibbot, UB, Xiai, Xpang, Galbet, and
more. Quinn Robot is Alibaba trying to
become the AI brain that connects all of
that hardware into one serious robotics
ecosystem. That's everything for this
one. The pace across all of this is
moving faster than most people were
predicting, and it's only going to get
more interesting from here. Thanks for
watching, and I'll catch you in the next
one.