New Clone AI Robot With Artificial Muscles Shocks The World_ Torso 4

Infinity Robotic
4 followers
Clone Robotics is pushing humanoids closer to the human body with Torso 3’s artificial muscles and lifelike movement, while the redesigned Torso 4 targets real workplace tasks. Meanwhile RoboParty RP1 takes brutal balance tests and AGIBOT deploys 100 humanoids in real stores.
#machine #Robotics #robot
AI Generated·May contain errors
A robot with a human-like skeleton, artificial muscles and tendons,
just moved its whole body like a living thing.
And the company that built it is already calling part of that hardware legacy.
That company is Clone Robotics, the robot is Torso 3,
and the thing replacing it is a fully redesigned machine called Torso 4
that's supposed to launch in November.
So, the new footage first.
Torso 3 is suspended from above while a human operator controls it remotely,
and the whole body moves together.
Shoulders, arms, the torso itself, hips, legs and hands,
all at once, all through these really wide motions.
And it's honestly a little uncomfortable to watch,
because almost nothing about the way it moves looks mechanical.
The reason it feels so strange to watch comes down to how Clone builds these things.
Most humanoids you've seen, like Optimus, Figure or Unitrix robots,
are built around electric motors and gearboxes sitting right at the joints.
Clone went a totally different route with what it calls a musculoskeletal approach.
At Clone, we've done the hard work of developing these muscle fibers.
There's a skeleton shaped like a human one,
artificial muscles attached around those bones,
and the whole thing runs on fluid control.
So when a muscle contracts, it pulls on the bone kind of the way yours does,
and your brain recognizes that movement instantly.
It moves like a body.
Which is probably why the comments went absolutely wild.
People called it Westworld.
Someone said Vekna.
A bunch of people just wrote Creepy or Please Stop.
And one person said that's exactly what they look like when they try to stretch.
A Russian commenter wrote that 2029 is getting closer.
Someone in Portuguese said this thing is going to cause a huge problem.
And another said this one is going to be a real humanoid.
My favorite, though, was an Italian commenter asking
how it's been hanging like that for months and still hasn't learned to walk.
And honestly, that joke kind of lands.
Because the rig is there for a reason.
This is teleoperated full body motion with no autonomous walking
and no balancing without support.
There's real progress under the hood, though.
Clone moved its valves in-house.
The valves are what control the fluid going into each muscle.
And this change brings more of that muscle control hardware inside the robot itself.
The account that shared the footage described Torso 3 as
And yeah, that's exactly it.
But Clone is already moving past Torso 3.
In a follow-up post, the company said that over the past six months,
the team redesigned the entire system from the ground up.
The priority this time was making the robot easier to simulate and easier to control
while still aiming to keep what Clone calls human-level capabilities.
That makes a lot of sense.
Since most robots now learn in simulation first,
practicing on a virtual copy before touching real hardware,
and a tangle of muscles,
tendons and pressurized fluid is really hard to model on a computer.
All of that redesign work ends up in Torso 4, launching in November.
And the target is very specific.
Clone wants to automate stationary bimanual tasks for enterprises.
Bimanual just means two-handed.
So this is work done in one spot with both arms.
Like someone standing at a workstation all shift.
Clone has shown off stuff before,
like really fast hand movements and a neural controller driving them.
But this announcement leans hard into simulation,
controllability, and an actual platform for workplace tasks.
It's Clone starting to sound like a company with something to sell.
Now, the hands?
Because honestly, that's where my eyes kept going.
On October 1, Clone posted a close-up demo with a very plain caption,
Teleoperation of the Hand of Torso 3,
and it pulled in around 850 likes.
It's a five-finger hand where artificial muscles pull on tendons,
all laid out the way human anatomy is.
In a follow-up, Clone said this hand was made back in April 2026,
and that it's the last of its legacy hand designs.
One upgrade on Torso 3 is joint angle feedback in the hands.
Finger movement is one of the hardest parts of controlling a humanoid.
With muscles and tendons, you can tell a muscle to pull,
but that alone doesn't tell you exactly where the finger ended up.
Now the controller gets direct feedback on where each joint actually moved
after the muscles were commanded,
which is exactly what you need for fine finger control.
Boston Dynamics revealed new hands for Atlas on the very same day,
just hours apart,
and Humanoids Daily summed up the contrast perfectly.
Boston Dynamics built its hand like a machine,
with four fingers, 13 degrees of freedom,
meaning 13 separate ways the hand can move,
a motor inside every joint, no tendons, and no pinky.
Clone built its hand like a body,
with five fingers pulled by artificial muscles and tendons.
Boston Dynamics dropped that fifth finger on purpose,
to keep cost, size, and complexity down,
and to cut the number of things that can fail.
Five-finger hands like Clone's go after the most natural range of grips possible,
basically anything a human hand can grab.
Two totally different philosophies,
and yet both companies are now pushing the exact same priority,
which is making their hardware easier to simulate and control
on the weight of hands that are actually useful.
Still, the hand demo is an older design under direct human remote control,
so it shows the mechanics of the muscle and tendon drive,
and says nothing about autonomy or about what Torso 4 can do.
Clone also hasn't shared any performance numbers,
any specific tasks Torso 4 has already completed,
or how much human help it needs while working.
Automating any stationary two-arm task is a product goal at this point,
and nobody has seen it demonstrated.
November should finally show which tasks Torso 4 handles on its own,
how often a person still has to step in,
and whether a robot built like a human body can hold up in regular industrial work.
And honestly, I'm curious,
would you feel comfortable standing next to this thing?
Let me know in the comments.
Meanwhile, at IROS 2026 in Pittsburgh,
one of the biggest robotics research conferences out there,
a company called Roboparty,
did pretty much the opposite of hanging its robot from a rig.
It let people kick it.
On September 28, Roboparty unveiled RP1,
also called Roboto 01,
and that was the robot's global debut.
At the booth, they ran a demo literally called Kick Me,
where visitors could walk up and push or kick the robot,
and RP1 would quickly adjust its posture and get its balance back.
That's a bold move in a room full of roboticists,
since it tests dynamic stability,
body performance,
and system-level robustness all at once.
What makes that recovery possible is PartyOS,
Roboparty's open R&D foundation for humanoid robots.
Inside it, there's a training framework called UFO
that discovers motor skills
through unsupervised reinforcement learning.
In normal words,
the robot works out its own movement skills through trial and error
without anyone handing it pre-recorded motion trajectories to copy.
That covers switching from one skill to another,
recovering when someone shoves it,
and getting back up after a fall.
On the hardware side,
RP1 puts out up to 160 newton meters of peak joint torque,
which is basically how hard its joints can twist
running on Roboparty's own in-house actuator modules called Romomo,
plus a real-time motion control system.
The company is building the body,
the actuators,
the low-level control,
and the AI algorithms as one system,
so everything gets validated together.
The bigger story with this one is the open-source part.
RP1 builds on RPO,
short for Roboto Origin,
a fully open-source humanoid project Roboparty
launched in January 2026
that's picked up more than 2,500 stars on GitHub.
It carries over RPO's already-validated robot body and workflow,
the Romomo actuators,
and a dexterous hand called the RP Hand.
Roboparty is calling RP1 the world's first high-performance,
full-stack,
open-source,
bipedal humanoid
made for researchers,
educators,
robotics developers,
and embodied AI teams.
And full-stack, in their words,
means a lot more than releasing mechanical files.
The plan is to gradually open up the motion control systems,
simulation environments,
SDKs,
training tools,
and PartyOS itself
so developers can get involved at every level,
from hardware and low-level control
all the way up to training models and deploying them.
PartyOS is also evolving across humanoid locomotion,
perception and interaction,
whole-body manipulation,
and eventually autonomous, intelligent humanoids.
Founder Yi Huang put it pretty directly,
saying humanoid robots shouldn't be closed systems
that only a handful of teams can build,
while everyone else can basically just watch,
and that the goal is for RP1 to become an open foundation
developers can access,
modify,
train,
and keep improving.
The team behind it is stacked, too.
Core members studied at Tsinghua's Yao class
and Peking University's Zee class,
which are elite computer science and AI programs in China,
along with Harbin Institute of Technology,
Stanford,
Carnegie Mellon,
Zhejiang University,
and others.
Roboparty already has a global community
of more than 5,000 developers.
It'll share more details on RP1's open-source roadmap this month,
with broader software and hardware releases
and a mass-production program following later in Q4.
Then there's a Guybot,
which is a lot less dramatic than robots getting kicked,
but it's happening in real stores with real customers.
AGI BOT says it has deployed 100X,
two humanoids across 100 stores run by ASD,
a cookware and kitchenware company in China.
The robots are out on actual sales floors,
dressed in ASD aprons and chef's hats,
doing customer-facing work right alongside human sales associates.
The announcement went out on September 28.
X2 welcomes shoppers,
asks about stuff like household size,
cooking habits and budget,
then recommends cookware and explains the features.
It also helps guide people around the store.
AGI BOT described one interaction
where a family wanted a wok for Chinese stir-frying
with a budget under 1,000 yuan,
so roughly $140.
The robot recommended an option.
And as soon as the family asked about promotions,
it handed the conversation over to a human associate.
That handoff is basically the whole design.
Explaining what a pan does
relies on information that pretty much stays the same.
Completing a purchase needs current, store-specific details,
and somebody has to take responsibility for the transaction.
So the staff handle promotions, inventory information,
and the final sale,
while the robot handles the welcome and the recommendation.
To show all of this off,
AGI BOT ran a three-day live stream from September 28 to 30,
broadcasting from stores in Wenling, Shanghai, and Wuxi
for 10 hours a day from 10 a.m. to 8 p.m. China time.
And a long broadcast like that
is honestly way more useful than a short, polished demo.
Real shoppers interrupt, change their minds,
and ask things that have nothing to do with cookware
and how the robot deals with that.
And when it decides to call a colleague over,
matters just as much as nailing a smooth product pitch.
There's plenty the announcement doesn't prove, though.
10 hours on air doesn't mean the robot runs 10 hours on one battery,
and footage from a few selected stores
doesn't verify what's going on at all 100.
AGI BOT hasn't shared sales impact, intervention rates,
fleet-wide uptime, customer service measurements,
or the commercial terms of the deal.
The release mentions mobility and multimodal interaction,
but it doesn't explain what kind of remote support sits behind it.
So you can't assume every conversation across all those stores
happens without any human help.
This rollout comes right after AGI BOT's deployment at ChimeLong,
the theme park group,
where it announced more than 300 robots
working in entertainment, visitor assistance, and hospitality.
Before that, AGI BOT reported a 20,000 robot production milestone.
But that number covers several product families
and counts total robots manufactured,
which is a different thing from robots actually deployed or sold.
The ASD deal gives a much more specific picture
of where one particular model is working.
And both projects put robots where people bump into them
during normal everyday stuff
instead of only at robotics exhibitions.
For ASD, the real test is whether X2 actually makes shopping better
and helps the staff serve customers.
The numbers that would really tell the story are
how many conversations end in a genuinely helpful recommendation,
how much staff time goes into supporting the robots,
and whether all that customer interest turns into purchases or repeat visits.
A hundred stores is a solid place to gather exactly that kind of evidence.
For now, AGI BOT has shown the scale it's claiming
and how it splits the work between robots and people,
and the business results are still to come.
All right, that's it for this one.
Thanks for watching, and I'll see you in the next one.
