The first robot vacuum I ever owned was a Neato Robotics XV Signature Pro in 2015, and it couldn’t distinguish a mouse from a sock.
I learned this the hard way when I returned from a trip and found that it had run over a glue trap and suctioned up a dead mouse. Then it ran out of battery and simply sat there, festering, until I got home and gagged at the stench.
I threw the entire robot vacuum into the trash and didn’t have the courage to buy another for years. In 2018, I tried the Ecovacs Deebot N795. It was worlds better in terms of app support, cleaning, self-docking and charging. It was also quieter than the Neato, which sounded like a fighter jet taking off.
But it still suffered from a fundamental weakness afflicting robots: It was lousy at recognizing and avoiding objects. It bumped into furniture and chair legs, ran over socks and shoes, and slammed into anything unfortunate enough to be on the floor.
On the plus side, it never puréed a mouse, but the cleaning patterns seemed random and ineffective. Letting it run unsupervised meant I would come back to find it either safely docked in a clean home or tangled in cables and out of battery.
As CNET’s expert for testing robot vacuums, I’ve seen the tremendous progress since those early days. They feature navigation technology like lidar, which used to be found only in cars and boast advanced sensors and AI training models that can recognize objects. They also mop, self-empty and self-clean. Certain robot vacuums even come with gripper arms that can pick up items scattered about your home, and lift systems that can boost them up stairs.
In short: Robot vacuums have gotten very good in an astoundingly short period.
So now we’re ready for the humanoid robots we’ve been expecting since forever, right?
We’ve been imagining an ultimate robot companion that navigates the home, does chores and acts as a family confidant, and robot companies are making it all seem tantalizingly close. Humanoids folding laundry, walking confidently and looking ready to fit in were the talk of CES, the massive consumer electronics show, earlier this year. There is no shortage of videos online showing off the potential of robot helpers.
Consider Neo from 1X Technologies, the most human-looking of all the robots being touted today, and at $20,000 to preorder, one of the most expensive. It’s a true bipedal robot with functional legs, a fabric-covered body and a set of tendon-driven five-fingered hands, with demo videos showing it knocking out chores around the home, including serving coffee, vacuuming and cleaning windows.
The promo videos are slick, and the preorders and 2026 launch date make it seem like humanoid robots are ready to ship.
But that’s not the case. What we’re seeing, from 1X and others, is still early-stage prototypes and well-polished “humanoid theater.”
The challenge isn’t just in physical design, engineering and hardware reliability; it hinges on establishing training data and developing large language models to run a physical AI machine so that a robot can coexist with humans in a living space with clutter, pets, kids and furniture.
Several experts I spoke to argue that humanoids are a flawed design and that a true, practical application of consumer robotics already exists in single-use models like robot vacuums.
But others, like Andrew Kang, CEO of Robostrategy, which invests in robotics companies, say humanoid robots are progressing along the same evolutionary path followed by now-ubiquitous devices like smartphones.
“Just as the iPhone consolidated your camera, watch, alarm clock and calendar into one device, a humanoid will eventually absorb the work of a vacuum, a window cleaner, a lawnmower and a physical care assistant,” says Kang.
Maybe someday. Right now, there are more fundamental issues to solve.
‘Preorder now’ doesn’t mean it’s coming to market
Single-task robots like robot vacuums and pool cleaners have become exceptionally efficient, but a humanoid must merge physical dexterity with contextual intelligence. That’s going to take a massive leap forward.
Some companies say they’re on the cusp of delivering humanoids to households. Neo has been available for preorder since late last year for $20,000 or through a $499-per-month subscription with a $200 deposit. The company says it will start shipping Neo this year. Even if it makes that date, the robots will still likely rely heavily on remote operation by actual humans.
Then there’s Isaac 1, a laundry-folding robot produced by Weave Robotics. It is also available for preorder for $7,999 upfront or $449 per month, with an optional $99 monthly subscription for “future capability updates.” The company says it will begin delivering the model in California starting in the fall, with broader US availability in 2027.
Isaac 1 has a humanoid torso, a wheeled base for stability and two arm grippers, which the company says work just as well for folding laundry as fingers do. Demo videos show it sorting items, folding laundry and carrying baskets, though it moves notably slower than Neo.
Hello Robot’s Stretch 4 is another humanoid take, and the most restrained. Stretch 4’s wheeled base has more in common with a robot vacuum than a human. A telescoping pole with sensors gives it height and houses forceps-like pincers.
The Stretch 4 is intended for assistive elder care in homes. It’s priced at $29,950, which Hello Robot CEO Adam Edsinger tells me is the cost to sell to research labs and covers R&D and manufacturing support. It’s already in more than 300 facilities, but it’s a basic model for specific use cases only.
“It’s easy to overestimate what’s actually happening,” Edsinger says. “We really want to infer from these [demo] videos that it can do all the other things that we imagine. We see it folding laundry. Well, it must be able to load the dishwasher. But practically, we’re really not there yet.”
There are other humanoid robots, like the $13,500 Unitree G1, Apptronik’s Apollo 2 and Boston Dynamics’ Atlas, which are intended for research, industrial and commercial purposes rather than consumer use. All three models are distinctly humanoid, with bipedal legs, a head, a face and hands with fingers.
Other startups are aiming to bridge the gap between commercial and consumer use. Tau Robotics plans to deploy humanoid robot cleaning services in San Francisco, but it’s in the very early stages of rollout with restricted availability.
Zooey Liao/CNET/1X Technologies/Weave RoboticsMerging the physical body with general intelligence is the biggest challenge
Robots will need something closer to true intelligence, not just a collection of training routines. They’ll need to understand not just the task they’re performing, like watering a plant, but also to generalize it to complete other tasks, like filling a pet’s food dish.
It won’t be easy.
“If a vacuum robot can still be baffled by a Lego brick due to limited edge intelligence, a humanoid robot faces a much higher computing challenge,” says Yunlin Kong, president of overseas business at Narwal. “Until a robot can generalize and fully comprehend its environment on the floor, adding a humanoid body is simply putting a ‘form without a soul.’”
Robot vacuums are engineered to be reliable and repetitive — people largely expect their floors to be cleaned the same way every time. They generally don’t require advanced machine learning or large language models to do their job well.
The LLMs underlying AI chatbots like ChatGPT, Gemini and Copilot, are largely trained on words so they can respond to your queries and predict what you’ll ask or do next. Complex robots that perform a wide range of tasks that vary from home to home will require world models — simulations created from video data — to understand what will happen and predict how objects and people move through physical space.
Rodney Brooks, professor of robotics at MIT and co-founder of iRobot, has a blunter assessment.
“There ain’t going to be no humanoids in no one’s home anytime soon,” he tells me. “This whole humanoid thing is total fantasy.”
Robots have gotten better, but they’re not flawless
For years, industrial robots — typically just robotic arms — have been safely bolted to concrete floors and isolated behind protective cages. There are more mobile versions in warehouses that can operate more freely in structured environments to lift and move items.
But a human-size robot roaming around a house, especially one with kids or pets, is a different, and potentially dangerous, proposition.
Yi Guo, a professor of electrical and computer engineering at Stevens Institute of Technology, who researches human-robot interactions, gets nervous when she hears that startups are trying to put a heavy, lumbering machine in homes or eldercare facilities.
“Physical safety is a first thing,” says Guo. “I’m not going to test that, even in the senior center, before I know the capability of this humanoid robot.”
Nicholas Nadeau, co-founder and CTO of Onix, a company that builds personalized AI models trained for health and wellness, points out that you’re trading absolute predictability for a system that might work brilliantly one moment and then fail to even turn on for the next five days.
Franklin Tanner, vice president of robotics and physical AI at Innodata, a data engineering company that trains AI models, argues that this physical threat is severely underappreciated. For a robot vacuum, the fail state is that it simply stops, gets stuck or runs out of battery.
Imagine a humanoid robot as a small refrigerator trundling around your home on two unsteady legs.
“A modern humanoid carries real mass and torque,” Nadeau says. “These are meter- to 2-meter-tall systems at anywhere from 30 to 60 kilograms — up to 100 pounds — and they move at a walking pace. That’s a lot of momentum and energy if anything goes wrong.”
Seth Sternberg, CEO of Honor Technology, a health tech company focused on aging, points out that the physical dangers of robotics are at odds with the desire to use them for eldercare, interacting with humans who are often in a fragile state.
The physical danger isn’t just about weight. A bipedal machine needs a great deal of physics to stay upright. If it begins to lose its balance, even its safety response could be a hazard.
“Just think about the amount of force it takes to balance, the energy required at that hip joint to support itself,” Edsinger says. “If that leg had to kick itself out because it started to trip, you’re effectively swinging a metal baseball bat around your kitchen.”
A laptop or smartphone crash might give you a frozen screen or an unresponsive keyboard. A robot system failure could cause a more literal crash.
Robot makers may not be fully accounting for that. Nadeau says that in most cases, they have a software “cheat” to fail gracefully using a line of code on the back end that tells the robot to duck out and hide in a corner, avoiding engagement with the situation. He also warns that this isn’t a complete solution.
“If a robot experiences a software or hardware failure while carrying a pot of boiling water, there is currently no industry-wide consensus on how to make it fail gracefully,” Nadeau says.
Robots with a heavy wheeled base or a single telescoping arm might be a safer alternative to fully humanoid designs.
Physical danger is one thing; privacy is another
We’ve gotten used to living in a world of smart speakers and connected home hubs. But robots roaming bathrooms and bedrooms represent a significant escalation in privacy risk.
A doorbell camera may keep watch on your front door, but most are fixed cameras that watch only what you decide it should. It can’t suddenly enter your bedroom or bathroom. That’s not the case for a mobile robot, packed with cameras and sensors, and given free rein in your home.
“The more useful the device, the more the security has to match the trust we’re placing in it,” says Paul Pioselli, CEO of Solace, a personal cybersecurity service company. “The reason a home robot is worth $20,000 is the same reason it needs bank-grade security: It understands your home better than anything you’ve ever owned.”
What makes the privacy problem more serious is that many humanoids are supported by a remote human teleoperator to complete tasks and serve as an AI bridge for things the robot isn’t yet capable of.
Every company I spoke with acknowledged the risk but also emphasized they’re taking measures to ensure privacy, using a mix of design elements and data security measures.
Some humanoids, like the Stretch 4, won’t ever act autonomously because they’re meant to be assistive robots. Others have privacy-minded components built in.
“Isaac 1 has physical shutters that will actually cover its cameras so that you can know when it’s not working, which is an incredibly important part of designing for people’s most intimate spaces,” says Evan Wineland, co-founder of Weave Robotics.
Your home isn’t a factory, and that’s the problem.
Robots already live and work in many homes, in the form of vacuums, lawn mowers and pool cleaners. In the US, 22% of households have a robot vacuum, according to Euromonitor estimates.
But moving from a floor cleaner to a humanoid exposes a key problem: Homes have nearly zero standardization. An apartment in New York City is a completely different environment from a sprawling ranch in the Sunbelt.
“There are a number of reasons why we have decided to start in the industrial market first,” says Aya Durbin, director of product for Atlas at Boston Dynamics. “The industrial market is where you find jobs that are dangerous, ergonomically unfriendly and understaffed. The return on investment and impact of deploying humanoids into these roles is obvious.”
Industrial robots excel in automotive assembly lines because tasks are highly repetitive and the environments are rigidly structured. While a home has some structure — most have basics like a kitchen and bathroom — the layout, size, location and placement of appliances and objects can vary hugely.
“A home is one of the least predictable environments an AI system can enter,” says Tanner. “Each residence is uniquely configured. Furniture is moved around. Light levels vary throughout the day. Humans place items in random locations. Pets continuously move.”
Zooey Liao/CNET/Getty ImagesThe domestic environment is also uniquely hostile to delicate machinery. Ivo Marocco, vice president and head of UX at Renesas Electronics, a Japanese semiconductor manufacturer, points out that a human home is a filthy, chaotic ecosystem full of pet hair, moisture and dust.
The real bottleneck isn’t just algorithmic “smarts,” Marocco argues, but rather physical durability. Camera sensors fail when covered in household grime, prompting the industry to adopt rugged, more advanced sensors from the electric vehicle market.
More advanced robots require more advanced sensors, since they’re not just using vision to navigate and recognize objects; they also need a sense of touch to know how much force to use.
Niko Reuss, CEO of Freudenberg Technology Innovation, a company that supplies key components for robots, points out that the sheer volume of household dust poses a silent, existential threat to on-board processors. Dust is the real enemy of sensitive electronics, with dust on chips and mainboards causing thermal management problems if the robot isn’t sealed well enough.
“Startups chase capabilities first, but the robot itself will not hold up without solving these unsexy, typical issues,” Reuss says.
Should home robots even be designed to resemble humans?
Humanoid robot design makes sense at first glance: They work in all the same places as humans, which is the design reasoning behind Neo and other models that adopt a human form factor with limbs and hands with fingers.
“We believe that the world should be built for humans and that any robots that help humans should do it in the same way humans do,” says Tom Sanocki, 1X Technologies’ vice president of engineering. “We don’t have to design our world around wheeled robots that may not be able to go upstairs.”
Several experts I spoke with acknowledged that there’s logic in building robots for the human world, rather than reshaping the world to accommodate robots with wheels or treads.
But the hardware hurdle is immense. According to Reuss, getting humanoids out of the lab requires overcoming three specific hardware challenges: integrating tiny spinning parts (micro-servos) for dexterous hands, mass-producing durable joints that can handle heavy loads in a lightweight design and vastly improving battery endurance.
iRobot Chief Engineer Adam Pope argues that the future of chores will look more like the appliances we already trust, like dishwashers and washing machines. But why are companies trying to force a general-purpose, bipedal body to do jobs that appliances already do perfectly?
A dishwasher can clean the dishes effectively every single time. It’s hardly the best use of a humanoid’s expensive, advanced technology to handwash the dishes, since it’ll never do it as well as a dedicated appliance, which is why the demos show them loading it.
This is evident in many slick demo videos; even in a heavily curated and optimized environment, humanoid robots just aren’t very fast at these tasks. At CES 2026, I watched Switchbot’s Onero H1 slowly and inefficiently load three items of clothing, taking a full 2 minutes to complete this minor task.
Humanoids may be labor-saving, but they’re not time-saving.
Hands versus grippers
Every company that’s making a humanoid robot has to make choices. One key decision is between grippers and human-like hands with multiple fingers. Both come with trade-offs: Grippers are cheaper and easier to implement, but full human hands are better for fine manipulation.
Manufacturers and analysts are divided on whether a five-fingered hand is even necessary for a humanoid robot. Long Wang, assistant professor at the Stevens Institute of Technology Department of Mechanical Engineering, suggests a bird’s claw with three fingers is better, because it’s easier to implement and can still handle tools like scalpels.
I saw this in action at the robotics lab at Stevens Institute of Technology, when Wang demonstrated a medical and surgical robotic arm he’s developing with students for the US Army. While operated by a student, the robotic arm wielded a scalpel to make an incision and performed compressions with an oxygen mask on a CPR dummy, though it still needed to be handed the tools.
Some companies I spoke with are using parallel grippers rather than five-digit hands, such as Isaac 1.
“We already fold 1,000 pounds of laundry every single week using parallel grippers,” says Wineland. He believes that the full hand is more of a nice-to-have and that many of the key chores can be done with the grippers.
But a benefit of the five-fingered hand is that it enables easier transfer of skills from humans to robots.
Five-fingered hands also fit well with our homes, which are designed for our proportions and capabilities, and functions like opening drawers and turning doorknobs.
Robot hands require high dexterity, which means they need to make microscopic movements, sense physical objects and use more advanced algorithms to determine how much force to use for gripping and lifting. All three things increase costs and make durability a challenge.
Legs versus wheels
There’s a similar debate about legs or wheels, with many of the same trade-offs as five-fingered hands and grippers.
The Stretch 4 and Isaac 1 use a wheel-based design for better balance and stability.
Part of the benefit also comes down to ease of implementation. Sunseeker CEO Terry Ma tells me that it’s simply a pragmatic choice to make with these early models. The wheel is better designed for robot models currently than legs; it’s naturally balanced, doesn’t require huge processing power or expensive hardware, and doesn’t need to be as concerned with latency for movements.
One of the best designs for the foreseeable future may be what’s already used in commercial and medical applications – a wheeled robot with a manipulator arm attached, giving it some ability to manipulate objects while navigating relatively easily.
But the skeptics don’t deny that bipedal legs are useful and likely necessary to give a true generalist robot mobility in cluttered home environments and in spaces that aren’t ADA accessible.
“We don’t have anything against legs in the fullness of time, but wheeled mobility is objectively safer in the short term, and simpler to manufacture, simpler to maintain,” says Weave’s Wineland.
Physical AI needs to vastly improve for humanoids to have a brain
Beyond marrying design and emotional intelligence, what humanoid robots currently lack is complex physical AI capabilities, meaning a real understanding of objects and how they behave in the world when subject to physics.
With an LLM, artificial intelligence is essentially reading a book on learning to ride a bike; physical AI models would teach a robot to actually ride.
The gap between learning the action and performing it, while accounting for all the different possible outcomes, is a huge roadblock. Humans can load a dishwasher without getting confused when a glass is a different color or a plate is a different shape. AI models lack the ability to generalize their experiences.
At Stevens Institute of Technology’s robotics lab, Guo demonstrated a mobile robotic arm’s ability to grab and toss items into designated baskets without using vision. On the objects it had been trained for, the robot performed successfully — but when a half-full water bottle was swapped in with its training objects, the robot’s algorithm couldn’t account for the shifting center of gravity, resulting in botched throws.
Another robot’s purpose was to catch a ball thrown at it by calculating where it expected the ball to land, but out of dozens of tries, it succeeded once (and I think there’s a good chance that was just pure luck, since I aimed for the net while throwing the ball).
Guo tells me the robot will get better over time.
“One of the biggest bottlenecks to bringing humanoids into the home is real-world training data,” says Daniel Chu, chief product officer at Apptronik, the Texas-based robotics company behind the Apollo 2 humanoid robot. “Solving that accelerates everything else.”
There are two main ways to train an AI to perform physical tasks. One is physical capture through a huge amount of raw, real-world data of physical actions, including failures. This requires first-person headsets or human teleoperators using VR headsets to operate a robot remotely.
“We have customers that come to us and say, ‘We want a million hours or 2 million hours of data,’” says Tanner. “Collecting that much data is hard … especially the data that actually matters for these home-use robots.”
The captured data also needs to be painstakingly annotated to link objects with the robot’s policies. If there’s a video of a coffee mug, the mug needs to be labeled to show whether the mug is empty or full of hot liquid, so that the robot can decide its “action state,” meaning whether it should grab the mug carefully by the handle.
There is no shortage of companies working on this. Nvidia’s Cosmos 3 AI foundation model uses vision to reason over objects, real-world events and physics-based world simulations. Innodata hires people to record themselves performing household activities with a GoPro to help train its models, and the company is building a physical motion capture studio in New Jersey that will train robots using actors who interact with them.
Apptronik is working on Robot Park, a 90,000-square-foot data collection and training facility for humanoids to learn in and has a partnership with Google DeepMind to have the robots working, collecting data and improving in real environments on real tasks.
The other method of training robots is with world models — simulators that essentially construct an artificial world with a representation of physics.
“A video implicitly encodes physics — such as showing a ball dropping to the ground — allowing the world model to learn a highly accurate approximation of how the real world will react,” says Sanocki.
Sanocki tells me 1X is using a world model that generates images and builds actions from them.
“That approach has been proven to generalize much better so that we don’t have to give it every single possible situation,” he says. “We can test those in our simulator, test those in software without having access to a robot or having it in a real-world environment.”
No one is providing the full picture on cost or manufacturing complexity
Building humanoid robots is a ridiculously expensive undertaking. Nadeau says that anyone quoting a $10,000 consumer price is operating at a loss and ignoring how much they’ll cost over their life cycle. He says $50,000 to $100,000 is the range where he starts to believe companies when they cite their price.
But a humanoid robot that costs as much as a luxury car is likely to be unappealing in a market where consumers balk at a robot vacuum priced above $1,000.
The upfront cost of the robot is only part of it. The real problem is reparability – in that it basically doesn’t exist.
Nadeau’s company doesn’t have a skilled workforce ready to repair and operate the robots. Currently, its “very expensive engineers” are flown around the world to fix and hot-swap robots, and there may be as few as five people capable of this.
It may be some time before robots are in homes of anyone besides deep-pocketed early adopters.
“It will take a few years to scale to manufacturing,” Kang says. “I think that’s going to be the bottleneck that we’re going to see. So it might not be until, say, 2030 that we see tens of millions of robots across people’s homes.”
One big unknown is how the Trump administration’s ban on importing new robot models will play out.
Given the likelihood of constant troubleshooting, software updates and R&D, maybe a subscription model makes sense?
This model already exists for both Weave Robotics and Neo, at least on paper, and Reuss says it’s attractive because the hardware risk and repairability concerns are with the manufacturer. Instead of buying a robot for $10,000, you lease it, with the potential to trade it in for upgraded models.
It would be like renting or leasing a car.
“You get hot-swap robots because no one actually owns the robot,” Nadeau says. “If there’s a new hardware revision, you can just make sure that all robots in the deployed fleet are at the latest revisions and that people don’t keep using a robot that’s 10 years old and out of service.”
Humanoid robots will overtake commercial spaces first
The eventual arrival of our robotic housekeepers depends entirely on these two paths converging: The hardware engineers building the agile body and the AI developers building the intelligent brain that can learn and adapt on the fly.
The one field where robots are already in significant and widespread use is factory and warehouse work, and many of the companies actually deploying robots, like Boston Dynamics, are designing them for commercial, enterprise and industrial use. Sunseeker’s Ma believes that putting robots in pharmacies and convenience stores is a nearer-term target for humanoid deployment, too.
But we’re not going to see an overnight leap in robotics technology and capabilities. Brendan Englot, interim dean of the School of Computing and professor of mechanical engineering at the Stevens Institute of Technology, envisions it being a 20- to 25-year timeline; Elmer Morales, founder and CEO at KoderAI, a multi-agent AI coding platform for apps and websites, thinks it’ll be close to between five and 10 years.
Tanner sees the industry evolving in phases over the course of the next two decades:
Today: Limited-purpose robots (vacuums, mowers, pool cleaners).
Next 3-5 years: Basic fetch-and-carry robotic assistants that can perform some manipulation within a controlled environment.
5-10 years: Multifunctional assistants that need occasional human help to complete activities. (This seems to be the industry consensus.)
20+ years: Fully autonomous humanoids capable of robust domestic tasks and physical care.
“The hardware is further along than most people realize. The software generalization problem is being attacked from every angle simultaneously,” says Morales. “The products coming out the other side of this research wave are going to change how people live at home in ways that feel genuinely science fiction today.”
Maybe.
Brooks, the most dubious expert I spoke with, tells me that truly practical, domestic humanoid robotics is a “mass hallucination.”
“I had a humanoid company. I built humanoid robots at MIT. I’ve been involved with humanoid robots for over 30 years. It is not going to happen that they are useful in the home,” he says.
At the very least, we’re in a long, long cycle of demo videos that tantalize, a perpetual motion machine keeping the dream alive. But if not even proven single-task robots like vacuums and mowers are near full adoption, it’s unlikely you’ll be unboxing a humanoid in your home for a long time to come.
Visual Designer | Zooey Liao
Art Director | Jeffrey Hazelwood
Creative Director | Viva Tung
Video Host and Editor | JD Christison
Video Producers | Owen Poole, Dillon Lopez
Project Manager | Danielle Ramirez
Editors | Corinne Reichert, David Watsky
Director of Content | Jonathan Skillings
