A very serious guide to buying your own humanoid robot butler

Science fiction is strewn with humanoid robots, from bad-tempered Bender in Futurama to cunning Ava in Ex Machina. And it has long seemed like that’s the natural home for such robots – on the screen and in books. The idea of a walking, talking, functioning robot with two arms and two legs has appeared to be a distant dream.

But things are changing. Last year, machines ran, boxed and even played football at China’s World Humanoid Robot Games, albeit sometimes falling over in the process. Meanwhile, companies have been readying their own range of humanoids that promise to do something a bit more useful: help around the house. At the tail end of last year, robotics company 1X opened preorders for its NEO robot, with deliveries expected this year –  a crucial milestone. And with several other companies also now offering humanoid assistants for sale, we are finally living in a world where you can essentially buy yourself a robot butler. There are options out there for a few thousand pounds, less than the price of a small second-hand car.

That’s why I have decided to create what is probably the world’s first guide for those considering making a purchase – and those who just want to know what all this adds up to. I will unpack not just each model’s vital statistics, but also explore four vital questions any would-be robot owner should ask. One thing is for sure: buyers should beware – they might get a lot less and a lot more than they bargained for.

What’s on offer?

It is fair to say the humanoid robot business has a buzz about it. More than 40 new models were unveiled by companies worldwide in 2025 alone, 60 per cent of which were from China, according to data collated by Morgan Stanley. That’s partly thanks to the rise of powerful artificial intelligence systems that help the robots function – indeed, investor and tech analyst Cathie Wood has called humanoid robots the biggest opportunity for AI to enter our lives. OpenAI CEO Sam Altman said in an interview last year that, in the future, “you’ll be walking down the street, and there’ll be seven robots walking past you, doing things”. Cryptocurrency investor Andrew Kang has said that “2026 will be the year of humanoid robots”.

I have selected five of the most-publicised and affordable models and taken a careful look at their abilities (see the boxes throughout this article). These robots are all available for ordinary people to order, or their makers say they will be soon. And the prices no longer look like science fiction. Unitree’s entry-level humanoid R1 is advertised at a little over £4000, and while 1X may be pricing NEO at $20,000 to buy outright, it is a more manageable $499 per month as a subscription. Not cheap, but wealthy people looking to maximise their productivity by outsourcing the housework might be tempted. Before you click order, though, here are some questions to ask.

Free newsletter

Sign up to The Weekly

The best of New Scientist, including long-reads, culture, podcasts and news, each week.

Sign up to newsletter

Do you really want or need a humanoid robot?

At first glance, most of the designs on offer from major robot manufacturers tend to be much of a muchness: a body, a head, two arms and two legs. The variations tend to be in what’s known as the degrees of freedom, meaning the range of hinge-like movements in which their “joints” can move. 1X’s NEO has 22 degrees of freedom in each hand and 75 in total. The Figure 03 robot, an updated version of the Figure 02, which has already been put to work in one of BMW’s factories, can do things like tidy and restock your cupboards, but has only 30 degrees of freedom in total.

The fixation on human-shaped robots is partly practical. Our homes and workplaces are built around human bodies, from stairs and door handles to counter heights and dishwashers. But that doesn’t mean a humanoid is always the best tool for the job, says Jonathan Aitken at the University of Sheffield, UK. “Do they have to be humanoid?” he asks. “For me, the answer is, ‘No, I don’t think they do.’” He points out that, while such a design may be attractive and attention-grabbing, it isn’t always the most cost-effective, reliable or practical option available. “If I were going to design a cooking robot, it wouldn’t be a humanoid,” he says. Joanna Bryson at the Hertie School in Berlin, Germany, is on the same page. “It’s just really bizarre, the whole idea that you want something humanoid,” she says.

If all you want is clean floors, you can already buy a programmable robot vacuum cleaner. If you want objects picked up and put away, you might not need a bipedal android, but instead something closer to a mobile platform with a good arm. California-based company Sunday is offering something like that: a robot called Memo with a humanoid upper half on a wheeled platform. The firm is currently offering a select few people the chance to test the robot at no cost before it goes on sale more widely.

Among the most expensive components of a humanoid robot are the actuators, which convert electrical energy into precise joint motion. These can cost $500 to $1000 each, so simpler models with fewer complex foot- or hand-like appendages should be more affordable. “Take a vacuum cleaner and imagine it if it has an arm on it,” says Subramanian Ramamoorthy at the University of Edinburgh, UK. “That kind of robot is much more likely to enter our homes.”

What do you want your robot to do?

If you are dead set on splashing the cash on a humanoid robot, then Aitken advises thinking carefully about the tasks you have in mind for it. “The big thing for me is understanding what you want to use the humanoid robot for, because so much other stuff flows down after that,” he says. Two big factors that will determine what it can really help you with are its physical capabilities and how smart it is.

Let’s take the physical side first. Do you want help with laundry? Carrying groceries? Loading and unloading a dishwasher? Tidying clutter? Many of these tasks sound simple, until you try to precisely specify how they are done. One thing that can help work out what will and won’t be possible is to think about the robot’s practical abilities.

First up is payload, or the maximum amount your robot can carry. Standing still, the NEO can lift around 70 kilograms, or about as much as some adults weigh, but it can carry along only about 25 kg, the weight of a medium suitcase. This is more than many other models, which can move the odd plate, but can’t ferry around heavy bags or piles of laundry. “The payload is of the order of low single-digit kilograms per arm, which automatically limits the tasks,” says Aitken.

Another major limitation of the current crop of humanoid robots is their energy capacity. The battery on the NEO, for instance, can last around 4 hours on a fast charge. Gentle tasks like slow walking between rooms and basic tidying are possible on today’s batteries. Beyond that, things get more challenging. “As soon as I’m starting to ask for it to be more capable and able to do more of the types of tasks that we humans might do, then I’m going to be putting more of a strain on battery life,” says Aitken. Cheaper models also inevitably have less-capable batteries: the Unitree R1’s battery lasts only 1 hour.

Brawn means nothing for a humanoid robot, though, unless it has the brains to back it up. Robots are powered by AI models designed to help them take inputs – mostly video from built-in cameras – and plan their movements and actions. The most popular humanoid robot operating system, developed by Nvidia, is called the GR00T model and is designed to enable robots to interact smoothly with their environments. Nvidia CEO Jensen Huang has previously said that “the easiest robots to adapt into the world are humanoid robots because we built the world for us”.

Whatever the operating system, though, the hope is that it can be trained on data on how humans move through the world, so that the robots can mimic it. Vision-language models of the type being produced by Nvidia and others are designed to take inputs through the camera “eyes” of the robot, then translate those into prompts for how to move. World models also help by giving the robot an internal predictive representation of the environment it operates in and how it works. The models are used to simulate “what happens if I do X?”, so it can plan before acting.

Some humanoid-makers are looking to use GR00T for their operating system, but other companies are going their own way. Figure’s eponymous robots are powered by the firm’s Helix visual-language action model, which works a little like an AI large language model, but with visual and verbal inputs, while 1X combines its own proprietary world model and a vision-language model called Redwood.

But these systems aren’t perfect. Ramamoorthy points out that a high-profile demo of Tesla’s Optimus humanoid in 2024 was alleged to have been assisted by nearby humans wearing VR headsets and operating the robots. In later demonstrations, Tesla owner Elon Musk has been forthright in denying that this teleoperation is in play. Still, if you imagine you can now buy a plug-and-play android, Ramamoorthy has a reality check: “I think for the average everyday person, I would be really surprised if things started happening very soon.”

In part, that’s because the technology is still being developed – and those who adopt an early-stage bot are likely to be assisting in its evolution. It is similar to how self-driving vehicle trials are fed back to help improve their future operations. Those who have already played with some of the humanoids, like NEO, have highlighted their strengths and weaknesses: although some simple tasks could be done autonomously, many others required a teleoperator – a human looking remotely through the bot’s “eyes” – to guide its actions. “At the moment, I don’t know what I would do with (a humanoid robot),” says Ramamoorthy. “Because they’re not yet capable enough for doing the things that normal people are interested in.”

There are also privacy concerns. Unlike a laptop webcam, the sensors won’t be sitting still. “On a robot, those things will be active all of the time while it is operational,” says Aitken, meaning it could be walking and talking its way through your private life. He urges people to ask what happens to the data generated by any humanoid they buy. “Who owns that data? Where does it come from? Where does it go?” Bryson goes further: “People think they bought a friend, and they just bought a spy.”

Neither Figure nor 1X responded to emails seeking a response to criticism about the way humanoid robots more generally gather data, though humanoid robot firms say they use the data solely to improve performance.

Will your robot actually work at all – or last?

Assuming none of that bothers you, you also need to think about what the headline price of a robot truly gets you. Some of the robots, such as the Unitree R1, launched in July 2025 at £4400, may feel within reach for affluent people. But buy that, and you pretty much get only the actual device – it doesn’t come with a built-in AI brain to help the robot work, so only the technically able who could install something like that need apply.

In that sense, some of the cheaper robots on offer could be a false economy. In fact, it is probably best to think of humanoids not as a one-off purchase like a dishwasher, but more like a car or a games console – in other words, something that will require ongoing servicing and updates to stay in good working order. So, the monthly fee of $499 for NEO might be a better long-term bet than the one-off purchase price of $20,000. Get this wrong, and you could end up with a robot butler that doesn’t last – or even work at all.

How to fix computing’s AI energy problem: run everything backwards

Artificial intelligence wastes an extraordinary amount of energy – but running every computer calculation twice, first forwards and then backwards, could drastically curb that problem

What could go wrong?

Finally, don’t just think about the possible benefits of your putative new robot butler – there could be drawbacks, too. “Even if you have a task that you do 100 times and 99 per cent of the time you do it successfully, 1 per cent of the time you don’t,” says Aitken. “What can happen in that 1 per cent of the time?” he asks. “Because that’s actually the question.”

Put another way, how bad could it be if the robot made a mistake in your home? It seems impressive that Figure 02 models have been helping to assemble BMWs. But homes are harder places for robots to operate in than factories, with their predictable workflows and well-disciplined workers. In a home, you have children, pets, clutter and people walking around absent-mindedly. Aitken says his 18-month-old son likes to skitter around his family home. If his human parents can barely keep track of his movements, is there any real hope a humanoid robot can?