A humanoid robot nearly beat a human worker after ten hours of sorting packages. The gap was just 192 packages, and the human finished with sore hands while the robot kept going. That race shows how close these machines may be to useful everyday work, but it also raises a bigger question: which robots are actually doing jobs today? This video looks at seven humanoid robots and the human abilities they are trying to match, from working a full shift and following spoken instructions to using dexterous hands and operating in real workplaces. We compare robots from Figure, AgiBot, Apptronik, Boston Dynamics, Tesla and others to see which ones are still in demos, which are in pilot programs, and which have already found paid work. Watch to the end for a side-by-side ranking of which robot works most like a human and which one may reach the job first. Subscribe for more plain-spoken coverage of AI and robotics.
Chapters: 0:00 Robot vs human: the package race 1:13 AgiBot A3 Ultra: a full shift 3:13 Apptronik Apollo 2: instructions 5:04 Atlas: new four-finger hand 6:51 Tesla Optimus: learning by watching 8:27 Figure 03: 30 unseen homes 10:08 Agility Digit 5: safe beside people 12:05 1X NEO: in your home 13:40 Why human-shaped robots 14:42 The pattern and the comparison
In May 2026, one of Figure's humanoid
robots spent 10 hours sorting packages
next to a human worker. Live on stream.
The human, an intern named Aim, won by just 192 packages,
12,924 to 12,732. He averaged 2.79 seconds per package
and the robot averaged 2.83.
Afterwards, he reported sore hands
and forearms while the robot kept on working.
Figure's chief executive, Brett Adcock,
said it would be the last time a human will ever win.
Robots are getting that close to real human work.
Yet most of the ones that look ready are not yet doing the job.
Tesla does not plan to deliver Optimus robots
to other companies until the end of 2027.
A home robot promised for this year still has
no confirmed customer delivery.
Meanwhile, one plain-looking robot has held
a paid warehouse job for more than two years.
This video looks at seven robots.
Each one adds an ability every human worker has,
from lasting a full shift to living in someone's home.
At the end, we'll line them up and see which
works most like a human and which is actually on the job first.
The first ability is the most basic
The basic thing any employer expects.
Getting through the whole day,
anyone who has worked a front desk knows the hard part
is not one tricky question.
It is staying alert from the first visitor
of the morning to the last one at closing
time.
Agibot, one of China's biggest humanoid makers,
built its A3 Ultra for that kind of day.
The Shanghai company unveiled it
at the World Artificial Intelligence Conference
in July 2026. According to Aggiebot,
it runs for up to eight hours of combined operating time,
enough for a normal working day. When its power runs low,
staff can plug it in or swap its battery,
and the robot can also go and charge itself.
It was designed for jobs where people judge a worker within seconds,
greeting and guiding visitors, and fetching and caring for them.
The company also lists
hotel reception, retail help, and security patrols,
lidar, and several types of cameras.
Give it a view all around, and it stops when someone touches it.
Its body grew out of an
earlier Aggiebot robot built for stage performances,
which gives it balance and agility.
It stands 1.74 meters tall,
about 5 feet 8 inches, and weighs 60 kilograms.
At the conference itself,
Agibot said more than 30 of its robots were working across the venues,
guiding visitors and answering questions.
The A3 Ultra was named one of the show's gems of the exhibition,
the only embodied AI product on
that list. In late October, the company proposed
a way to measure robots the way people are measured
at work, by how many hours they work on their own
and how much they get done.
There are limits. Each arm carries about five kilograms,
around 11 pounds. So this is front
of house work rather than heavy lifting.
And no outside group has yet published data from its
real shifts. Lasting the day is the first step,
but a good worker also has to understand what
they're asked to do.
The best new hires are the ones you can simply talk to.
Tell them to put a box on the bottom shelf
and they work out the rest without a manual
apptronic, a robotics company from Austin,
Texas, built Apollo 2 to work that way.
Its understanding comes from Google DeepMind.
In July 2026, DeepMind released Gemini Robotics 2
and showed it running on Apollo 2.
Earlier versions mostly controlled an upper body
at a table. This one controls the whole body at once,
walking, crouching, and balancing while
While handling objects with five fingered hands.
In one demo, Apollo 2 was given a single instruction.
Put the watering can in the green bin on the bottom shelf.
Nobody programmed the steps.
The robot walked to a table, found the can, and picked it up.
Then it stepped over to the shelves, bent down,
and placed the can where it had been asked.
Other demos showed the same system tying a trash bag
and unscrewing a light bulb. These are small jobs, but each one starts
with understanding a request made in everyday words.
Apptronic is already testing Apollo in real workplaces.
Mercedes-Benz has run pilots with
Apollo robots since 2024, moving parts and doing quality checks.
In June 2026, Apptronic opened an
expanded robot park in Austin,
a training facility of about 90,000 square feet where fleets of robots
practice tasks and collect data through a mix of teleoperation
and autonomous work. Apollo 2 comes
He comes on two legs.
The catch is that this is still pilot work,
no customer has published results,
Apptronics' chief executive has described Apollo 2
as a platform for collecting data,
and you cannot buy one yet.
Understanding a job only helps if the robot has hands that can do it.
Try taping your little finger to the finger next to it for a whole day.
You would still be able to cook, type,
open doors, and carry your shopping.
That simple test shaped the newest hand for Boston Dynamics at Atlas.
Boston Dynamics, which is owned by Hyundai,
revealed the hand in October 2026.
It has four fingers instead of five.
It can hold and use tools with triggers, like drills,
turn small objects around inside its grip,
and recover a grasp that starts to slip.
In one demo, Atlas held two balls in one hand
and rolled them around each other.
Engineers measured this kind of skill in degrees of freedom,
the number of separate ways a hand can
move. The previous Atlas hand had seven.
The new one has 13, and it is about the size of a large
human hand. The missing finger was a deliberate choice.
Engineers spent a day with their little
finger tied to their ring finger.
And at the end of that experiment, the company says they agreed
no pinky. A fifth finger would have meant more motors,
more costs, and one more part that could
Those hands are already training for real jobs.
Since September, Atlas robots have been practicing
with real car parts
at Boston Dynamics' new application center inside Hyundai's factory
complex near Savannah, Georgia,
learning to sort parts and deliver them to the right place on the
line. Hyundai has said it plans to deploy tens
of thousands of these robots across its plants.
Plants still.
The first factory jobs are not planned until 2028
and trainers sometimes take remote control
of the robots to teach them new moves.
A capable hand is only half the story.
A good worker also has to keep learning new tasks.
Most people learn a new job by watching
someone who already knows it,
then trying it and practicing until it becomes second nature.
Tesla wants its Optimus robot to learn in a similar way.
On the company's earnings call in
October 2026, Elon Musk described the plan.
Optimus first watches people do a task and
then practices it on its own until it gets it right.
He said there is no suit a person can
climb into to steer the robot like a puppet,
so it has to learn each task itself. To make that work,
Tesla is building what it calls an Optimist Academy.
where many robots practice a wide range of jobs
and feed what they learn back into training.
Unlike its cars, Tesla does not yet have
a large fleet of robots collecting data,
so it has to build the robots first and let them practice.
That building has started.
Tesla says it has begun production of Optimus
On its high volume lines in Fremont, California.
Each robot needs about 10,000 unique parts,
which is why Musk expects the early ramp to be slow.
He has also said the hand, and especially the fingers,
is the hardest engineering problem his team faces,
which links straight back to Atlas.
There is still a long way to go.
Optimus Gen 3 has not been publicly unveiled.
There is no announced price.
Deliveries to outside companies are planned for around the
end of 2027, and Musk himself has asked people
to temper their expectations.
Learning a task is one thing.
Doing it somewhere you have never been is much harder. On your first
day cleaning someone else's home,
the bed is a different size and the towels are in an unfamiliar
cupboard. Most people still manage
because they already know how to do the chores.
Figure the California company behind
the package race wanted to see if its robot could do the
same.
The company rented 30 homes in the San Francisco Bay Area
that the robot had never entered
and collected no data in them beforehand.
The robot had three chores, making beds, folding towels,
and tidying living rooms, using whatever
beds, couches, and surfaces each home had.
Across 420 full task attempts, it completed 237.
The comparison behind that number matters.
Figure says the same type of model trained without
its library of human video succeeded only 9% of the time.
That library, which Figure calls index,
is made of clips of people doing real tasks at home and at work.
The hands help too.
Figura 3 has cameras in its palms and fingertip sensors that,
according to Figura, can detect
forces as small as about 3 grams, roughly the weight of a paperclip.
That helps with soft, floppy things like towels and sheets,
and this is the same robot model
that came within a few hundred packages of a person over a full shift.
The limits are clear.
In a new home, the robot still failed on nearly half of its attempts.
And while the homes were
new, the chores themselves have been learned elsewhere first.
Working in a strange place
is impressive, but a real workplace
also needs a robot that people can trust beside them all day.
Every warehouse has that one coworker
who sees you carrying something heavy
and steps out of the way without being asked. Nobody notices it.
But it is part of why people feel safe working together.
Agility Robotics built Digit5 to be that kind of co-worker.
The Oregon company unveiled Digit5 in September 2026
as a humanoid designed to work near people
without a safety cage. Most factory robots work behind fences
because they are not built to react
when a person gets too close.
Digit5 watches for people using AI and several kinds of sensors.
When someone steps into its space,
it can steer around them, stop, or sit down.
It signals what it is about to do with lights
and sounds and a separate safety
controller watches over its movements.
It is built for hard, repetitive work.
It can repeatedly lift up to 50 pounds and according to Agility,
it recharges in about 9
minutes. The company also replaced the backward bending legs
that made earlier versions look like
birds with a new forward-facing knee design.
What sets Digit apart is that it already has a job.
An earlier version has worked in a GXO warehouse
in Georgia under a paid agreement
since 2024, and it has moved more than 100,000 totes there.
In late October 2026,
GXO announced it would bring Digit to two more warehouses
in North Carolina and Tennessee starting in early 2027.
There are limits here too.
DigiFive reaches its first customers in 2027
and the work Digit has proven so far is simple
tote moving and a warehouse is a controlled environment.
The hardest workplace of all
is the one where people actually live.
Letting someone new into your home takes trust.
A cleaner or a caregiver sees your kitchen,
your bedroom, and how you live. The seventh robot wants
that job, and it was designed for the home from the start.
EO comes from 1X, an open AI-backed
company that builds its robots in California.
It stands about 1.65 meters tall, around 5 feet 5,
and weighs about 30 kilograms. Its body is soft.
With a fabric-like covering and no pinch points,
it runs at around 22 decibels, very quiet,
and when its battery runs low, it plugs itself in.
One X says NIO will handle chores like
fetching things, carrying groceries, and loading the dishwasher.
It talks through a built-in AI and remembers past conversations.
It costs $20,000, or buyers can choose a monthly plan instead.
And the company has said its first year
of production sold out within days of opening orders.
There is a twist.
For chores Neo has not learned yet.
Owners can book a session in which a 1X employee controls
the robot remotely through a VR headset.
The owner approves each session
and the robot's indicator lights show when someone is connected.
When the Wall Street Journal's Joanna Stern spent time
with an early unit, the robot was
tele-operated for most of what she saw.
That raises an obvious privacy question about
a stranger seeing inside your home.
There is also a timing question.
NIO was promised for homes this year,
but no customer delivery has been confirmed.
After seven robots and seven human abilities,
none of them can do everything a person does.
So why do companies keep building workers in our shape?
The first reason is practical.
Factories, warehouses, and homes were built around human
bodies from the height of shelves to the width of doors
and the shape of tools. Stairs, ladders,
and car seats all assume two legs and two hands.
A robot with a human shape can walk into those
places and start working without anyone rebuilding them.
The second reason is communication.
A worker you can simply talk to fits
into a team without anyone learning a new way to
give instructions.
The third reason is learning.
The companies in this video teach their robots by showing them people.
Figure trains on videos of humans doing chores.
Tesla wants optimists to watch before it practices.
Boston Dynamics kept the Atlas hand close enough
to ours that human demonstrations can
help train it.
A robot built like us can learn from us,
and there are billions of people to learn from.
Lying all seven robots up side by side and a pattern appears.
The more human-like the work
a robot shows, the more it still depends on demos,
tests, or a person helping from far away.
The A3 Ultra can last a day, but it does light service work.
Apollo 2 can understand a spoken
request, but it is still in pilots.
Atlas and Optimus both have impressive skills
and both are waiting on factory dates that are a year or more away.
Figure 03 can work in a home it has never seen,
but only a little more than half the time.
Neo aims for the most human job of all,
yet a remote person still steps in for the hard chores.
Digit breaks that pattern.
It looks the least impressive of the seven and does the simplest
work, yet it is the only one with years
of paid work in a real warehouse behind it.
So where does that leave the comparison?
On the evidence covered here, Figure 03 has shown
the most human-like range of work.
It came within 192 packages of a person over a 10-hour shift,
and it carried household skills into homes it had never seen.
Digit is the robot that is actually
on the job, and its cage-free successor
is due to reach customers in 2027.
No robot has yet combined both.
Which of these seven do you think
will be doing real work first?
Let me know in the comments.
Thanks for watching and hit the like and subscribe.