Here’s how we could finally build humanoid robots that do all our domestic chores
By Erico Guizzo & Randi Klett
Published 11/11/2024
This is a non-interactive version of our story “Where’s My Robot?” — it includes all text, images, and videos (note that videos have no audio).
This story is part of IEEE Spectrum’s “Reinventing Invention” special issue.
See the interactive version on our site by using a browser with JavaScript enabled →
A new generation of advanced AI-powered humanoid robots that includes Neo from 1X Technologies, Apollo from Apptronik, Figure 02 from Figure AI, Phoenix from Sanctuary AI, Optimus from Tesla, and Atlas from Boston Dynamics is bringing us closer to the home robot helpers of our sci-fi dreams. What breakthroughs in AI and robot hardware could finally turn an autonomous robot assistant that do all our domestic chores, like Rosie from The Jetsons, into reality?IEEE Spectrum
Humanoid robot Atlas HD can run and jump over obstacles.ATLAS HD (BOSTON DYNAMICS)
Some people worry robots could maybe get a little too advanced…
The latest version of Atlas can do push-ups; another humanoid, Nadia, practices its boxing skills under teleoperation; the robot G1 cracks a walnut with its fist; and two Atlas HD robots perform two backflips in a row.ATLAS/ATLAS HD (BOSTON DYNAMICS); NADIA (IHMC); G1 (UNITREE ROBOTICS)
…and try to take over like in sci-fi movies.
T-800 from The Terminator movie.“THE TERMINATOR” (1984)
The reality is robots still need a lot of improvement.
Before they can rise up, robots need to stop falling down.
Three robots collapse during the DARPA Robotics Challenge Finals; Atlas HD humanoids stumble and fall down on an obstacle course at Boston Dynamics.FLORIAN (TEAM VIGIR); HERCULES (TRAC LABS); JOHNNY 05 (TEAM HECTOR); ATLAS HD (BOSTON DYNAMICS)
So forget the evil robots.
Let’s talk about the helper robots we’ve always dreamed of.
Household humanoids could do tasks like preparing and serving meals.
Robots to help us with chores around the house.
Cleaning. Cooking. Organizing.
Household humanoids could also do cleaning tasks like dusting surfaces.
Even keeping us company.
Robin Williams plays the android Andrew in the movie “Bicentennial Man.”“BICENTENNIAL MAN” (1999)
These robots would give us something of immense value: more free time.
Rashida Jones plays an American woman living in Japan with a domestic robot called Sunny in the Apple TV+ show “Sunny.”“SUNNY” (2024)
They could also help aging individuals remain independent.
Frank Langella stars in the sci-fi comedy-drama “Robot & Frank,” which features a companion humanoid robot.“ROBOT & FRANK” (2012)
We were promised Rosie.
Rosie, the household robot helper from “The Jetsons.”“THE JETSONS” (1962-1963)
Instead we got Roomba.
Roomba is known for its disc-shaped design.ROOMBA (IROBOT)
So the question is…
Where’s My Robot?
The Robot Race Is On
First the good news: It’s an exciting time for robotics right now.
New humanoids are jumping into action at a pace never before seen.
The robot race is on. Check out some of the contenders.
Who’s Going to Build the Best Robot?
Apollo
Apptronik
Apptronik has worked on more than half a dozen humanoid robots over the past eight years, including NASA’s Valkyrie. Apollo is the culmination of all this experience and is designed for manufacturability. The company plans to release the robot commercially following a series of pilot projects to demonstrate its autonomous skills.Apptronik
Atlas
Boston Dynamics
Boston Dynamics is known for its highly agile robots. The company, acquired by Hyundai in 2020, has recently unveiled a new fully-electric Atlas humanoid. Unlike previous generations, it doesn’t rely on hydraulics and is lighter and more compact, while also being able to move and bend its body in ways that exceed human capabilities.Boston Dynamics
Digit
Agility Robotics
Created by Agility Robotics, Digit is most accurately described as “bipedal” rather than “humanoid.” It has two legs, but its legs look more like those of an ostrich rather than a human’s. This is a side effect of Agility’s design process, the goal of which was to maximize the efficiency and robustness of legged locomotion.Agility Robotics
Figure 02
Figure
Founded in 2022, Figure has a very experienced team, which traces its heritage back through the DARPA Robotics Challenge. By iterating very quickly through hardware prototypes, Figure expects to demonstrate commercial viability in real-world use cases, and scale from there to “integrate humanoids into the labor force.”Figure
GR-1
Fourier Intelligence
Singapore-based Fourier Intelligence is already mass-producing its GR-1 and GR-2 humanoids. The company’ background is in health-care robotics, and it sees potential applications in medical and rehabilitation contexts, with the robots also available to researchers seeking a humanoid development platform.Fourier Intelligence
H1
Unitree Robotics
Unitree, founded in China in 2016, is well known for its capable and low-cost quadruped robots. H1 was Unitree’s first humanoid, and the company has already introduced a new, smaller model, the G1. Both are designed to be affordable platforms for research, or for companies focused on developing software rather than hardware.Unitree Robotics
Neo
1X Technologies
1X’s soft, tendon-based robot Neo is designed to have very low inertia, with the goal of building a robot that’s safe for humans to be around. The robot will weigh just 30 kilograms, with a carrying capacity of up to 20 kg. 1X, backed by OpenAI, hopes that Neo will become “an all-purpose android assistant to your daily life.”1X Technologies
Optimus
Tesla
Tesla has a few unique advantages when it comes to building and deploying robots. The company has substantial experience in battery technology, as well as in advanced sensing, computing, and AI for mobile systems. And Tesla is potentially its own first customer for humanoids, finding work for them in its car factories.Tesla
Phoenix
Sanctuary AI
Sanctuary AI’s goal is to “create the world’s first humanlike intelligence in general-purpose robots.” The company has been collecting extensive amounts of data of humans teleoperating its robots through complex manipulation tasks. Sanctuary AI hopes to leverage that data to train its robots to perform those tasks autonomously.Sanctuary AI
New Humanoids, New Skills
These robots are learning a lot of new skills.
Neo can lift an object like a backpack and hand it to a person.NEO (1X TECHNOLOGIES)
They work hard.
Digit was put to work carrying bins in a warehouse.DIGIT (AGILITY ROBOTICS)
They play hard.
Optimus robots from Tesla display their dance moves.OPTIMI (TESLA)
And they are constantly trying to outdo each other.
Some believe that the current artificial intelligence boom will propel robots too.
AI could help robots learn tasks that have proved difficult for them like grasping arbitrary objects.OPEN X-EMBODIMENT COLLABORATION
AI could finally solve the hardest problems in robotics—such as performing complex and dexterous tasks.
This two-armed robot powered by AI created by Physical Intelligence can fold clothing autonomously.π0 (PHYSICAL INTELLIGENCE)
These problems have largely kept humanoids confined to labs as experiments.
Phoenix can grasp cups of different shapes, sizes, and materials.PHOENIX (SANCTUARY AI)
The recent wave of AI breakthroughs gave us ChatGPT.
Next, will it teach robots how to make you a sandwich?
AI could finally teach robots useful tasks — like making a sandwich.PHOENIX (SANCTUARY AI)
Now for the bad news about this fresh batch of bots.
They’re not headed straight for your home.
They’ll first be found in factories and warehouses.
Figure 02 robots working at a car factory.FIGURE 02 (FIGURE AI)
That’s because they are best suited for dirty, dull, and dangerous jobs—jobs you find in an industrial setting, not in our homes.
Humanoid robot Apollo can lift and carry boxes from one place to another.APOLLO (APPTRONIK)
These robots are still too complex, too costly, and too unsafe for home use.
You definitely don’t want one of them staggering in your kitchen.
Atlas HD accidentally crashes into objects and falls down.ATLAS HD (BOSTON DYNAMICS)
So in an era of promising breakthroughs in…autonomous vehicles, planetary exploration, quantum computers, reusable rockets, fusion reactors, drug discovery, and more…
Japanese automaker Honda developed the first prototypes in the 1980s.
It took decades and vast sums to turn Asimo into a skilled humanoid.
The legged robots in Honda’s E Series were the precursors to its full-body humanoids.E SERIES (HONDA)
The Evolution of Asimo
Honda’s humanoid efforts started with the E series of bipedal robots in the late 1980s, advanced through the 1990s with the P series, and culminated with its most advanced full-body humanoid, Asimo, in the 2000s.E SERIES, P SERIES, ASIMO (HONDA)
Why Build Robots That Resemble People?
In a 2011 demo, Asimo danced, walked over uneven terrain, and ran in circles.
It also climbed stairs, hopped on one foot, and kicked a ball.
Asimo was able to perform an impressive variety of tasks.ASIMO (HONDA)
But the flawless performance belied a more underwhelming reality:
Mori-san conceived this clever chart to convey his idea.
Let’s look first at an industrial robot arm.
Your reaction to its appearance is likely neither positive nor negative.
Now consider a cute humanoid robot.
It can elicit a more positive response because it begins to resemble people.
A human being produces a maximum positive response according to Mori’s chart.
That’s because evolution made humans really good at recognizing other humans.
But our reaction plunges into negative territory if we see a creepy robot.
If it’s not perfectly lifelike, it may resemble a human corpse or a zombie!
That creepy zone is the Uncanny Valley.
This illustration shows a simplified version of the original chart that Japanese roboticist Masahiro Mori presented in his 1970 article on the Uncanny Valley.
The Uncanny Valley is a conjecture, not scientific fact.
Still, roboticists find it helpful when exploring humanoid design.
Some think that ultrarealistic androids are best for interacting with people.
Others avoid lifelike robots—they don’t want to risk “falling” into the valley.
All right, let’s get out of here.
Robot Hardware and Robot Software Challenges
Let’s get back to the challenges of building humanoids.
It comes down to two things: hardware and software.
Aila uses 14 brushless DC motors for its arms, four linear DC motors for its torso, and two servo motors for the head, plus additional motors for its mobile base.AILA (DFKI ROBOTICS INNOVATION CENTER)
The first three—sensors, computers, and batteries—are capable enough.
They keep getting better thanks to Moore’s Law and new battery innovations.
Phoenix has multiple cameras in its head.PHOENIX (SANCTUARY AI)
But actuators have lagged behind.
No fundamental breakthroughs have improved them on a similar scale.
Actuators are typically cylindrical devices, with a metal frame and rotating shaft at its center.
The most advanced actuators consist of a DC motor coupled to a gearbox.
Inside there are also sensors and power electronics.
An advanced actuator packs many components, including a brushless DC motor, position encoders, and a reduction gear system.DYNAMIXEL-Y (ROBOTIS)
These actuators are powerful and accurate, thanks to precision-machined parts.
Maxon makes some of the most popular robot motors, with the company’s devices used in a variety of applications, including research humanoids, soccer-playing mobile robots, and Mars rovers.HEJ 90-48-140 (MAXON)
The challenge is that they are complex, bulky, and expensive.
And a humanoid may require 20 or more of them.
Every active degree of motion in a robot’s body requires an actuator.G1 (UNITREE ROBOTICS)
Some robot makers are trying to up their actuator game.
They are making compact actuators that are strong and robust against, er, impacts.
Getting kicked is part of H1’s balance and robustness testing.H1 (UNITREE ROBOTICS)
Others are testing direct-drive-type motors, which do away with gears.
Or cable-driven actuators, which are lighter and safer.
AMBIDEX is a cable-driven robot that can give you a high-five and won’t break your hand.AMBIDEX (NAVER LABS)
Other alternatives may emerge by seeking novel bioinspired designs.
One way or another, robot actuators are in need of a radical reinvention.
This hand uses thin cables to actuate its finger joints.HAND ARM SYSTEM (DLR)
Next, let’s look at the software side of things.
Vision, navigation, control—so many components need to work together.
Consider, in particular, robot learning algorithms.
The goal is learning a task—like opening a door—under varying conditions.
Eve is learning to open different kinds of doors.EVE (1X TECHNOLOGIES)
But what if your robot encounters a new type of door and it doesn’t know what to do?
We need robots to learn like we do—humans adapt, correct, and improvise.
Japanese robot HRP-2 suffered a malfunction, shutting down and falling backwards, while facing a closed door at the DARPA Robotics Challenge Finals.HRP-2 (AIST-NEDO)
Enter AI.
Robotics and artificial intelligence have long evolved hand in hand.
G1 learned this skill using AI and simulation.G1 (UNITREE ROBOTICS)
So it’s no surprise that the rise of generative AI has electrified the robotics industry.
The hope is that AI could finally unlock the power of robot learning.
Google Deepmind researchers and their collaborators are using AI to train robot arms to perform a variety of grasping and manipulation tasks.OPEN X-EMBODIMENT (GOOGLE DEEPMIND)
As with most AI systems, the first step is amassing lots of training data.
One method involves humans teleoperating robots to perform various tasks.
Humans performing tasks while equipped with VR gear and wearable sensors generate AI-training data that can be used to teach robots to perform those tasks autonomously.OPTIMUS (TESLA)
Like making a sandwich.
Phoenix robot, controlled by a human teleoperator, attempts to assemble a sandwich.PHOENIX (SANCTUARY AI)
Next you use the data to train an AI model, and then test it on a real robot.
Training can also happen in a virtual environment with simulated robots.
Virtual robots learn to walk and climb in AI-powered simulations as part of Nvidia’s Project Gr00t, which focuses on AI for robotics.PROJECT GR00T (NVIDIA)
It’s unclear if AI can finally “solve” robot learning.
For one thing, training robots one task at a time is a very slow process.
A human operator remote controls the Apollo humanoid to make juice.APOLLO (APPTRONIK)
One way to speed things up is to develop a general-purpose AI model for robots.
A single model would let you perform a variety of tasks on a variety of robots.
A general-purpose robot model, trained on a large robot interaction dataset, would be able to control many different robots to perform many different tasks.π0 (PHYSICAL INTELLIGENCE)
Getting this data and training the models will take a lot of time and money.
But if it works, AI robots may finally be ready for the real world.
Neo unloads glasses from a dishwasher; Figure 01 operates a coffeemaker; and Digit removes a tray from a kitchen oven.NEO (1X TECHNOLOGIES); FIGURE 01 (FIGURE); DIGIT (AGILITY ROBOTICS)
Are the Robots Coming or What?
Where does that leave us? Are the robots coming or what?
We don’t know for sure how robots will evolve.
So let’s explore three possible futures.
Eve stands next to 1X Technologies founder and CEO Bernt Bornich.EVE (1X TECHNOLOGIES)
Each assumes a different level of technological progress.
And each imagines how robots may become part of our lives.
Figure 02 contemplates its own advanced hands.FIGURE 02 (FIGURE)
FUTURE 1: Small advances in robot hardware and software
Actuators are still chunky and pricey, and robot learning algorithms remain stuck.
Amazon’s Astro is a home robot that can drive around, play music, and record video.ASTRO (AMAZON)
Sorry, don’t expect Rosie to show up at your home.
But you can get a little robot buddy like this one from Amazon.
Astro can create a 3D map of your home and navigate autonomously.ASTRO (AMAZON)
It roams around your house taking photos of your pet.
You can remotely connect to your Astro to check in on your pets.ASTRO (AMAZON)
As for humanoids, they find employment in factories and warehouses.
Digit robots never get tired of moving bins in a warehouse.DIGIT (AGILITY ROBOTICS)
Except for Atlas, who becomes the first YouTube robo-influencer.
This robot’s dance repertoire is quite large.ATLAS HD (BOSTON DYNAMICS)
FUTURE 2: Moderate advances in robot hardware and software
New designs improve actuators, and AI makes robot learning more practical.
Stretch’s arm stretches so the robot can reach objects on tables, shelves, and countertops.STRETCH (HELLO ROBOT)
We’re moving out of lab-demo territory, and robots can finally do useful tasks.
These aren’t the general-purpose humanoids from the movies, though.
Stretch uses a simple robotic gripper to grasp household items.STRETCH (HELLO ROBOT)
These handy bots are known as mobile manipulators.
They do your laundry and clean your kitchen.
This two-armed mobile robot from Physical Intelligence can empty a dryer and fold clothing fully autonomously; the Stretch robot from Hello Robot, under teleoperation, can wipe a kitchen counter and load a dishwasher.π0 (PHYSICAL INTELLIGENCE); STRETCH (HELLO ROBOT)
They can even take care of your pets for you.
What more would you want? Pfft, a sandwich?!
Stretch, under teleoperation, feeds a treat to a dog, and opens the door to let another dog in.STRETCH (HELLO ROBOT)
And the humanoids? They’re still toiling in factories and warehouses.
Humanoid robot Atlas moves plastic engine covers from one storage cart to another fully autonomously.ATLAS (BOSTON DYNAMICS)
FUTURE 3: Massive advances in robot hardware and software
Big breakthroughs boost actuation and robot learning beyond expectations.
Future humanoid robots will need fine motor skills and adaptability to handle cooking utensils, pots and pans, and ingredients.
At last the future we’ve been waiting for.
The first generation of robot helpers is now ready for our homes.
Tomorrow’s robot helpers will also need to adapt to various household settings, such as different layouts, furniture, and appliances.
Yikes, these droids are pricey! You need a robot lease and robot insurance.
(Who knew robots would be just like cars!)
These robots will have to be extremely safe when operating and navigating in our homes, especially when performing tasks involving heat, chemicals, or sharp objects.
But what can we say?
Our new AI robots are the best. They do the chores. They take care of things.
Sit back and relax while Rosie makes you a sandwich.
If the right breakthroughs in AI and robot hardware materialize, a friendly robot helper like Rosie from The Jetsons could one day become part of our lives, helping out with our chores, and more.“THE JETSONS” (1962-1963)
