Elon Musk said Tesla’s Optimus humanoid could be on sale to the public by the end of 2027. Meanwhile, Google DeepMind’s most advanced robot brain still fails when it’s asked to survey a kitchen and put the ingredients for a mushroom risotto into a basket.
Those two facts are what’s happening in robotics right now. The pitch is ahead of the hardware, and the hardware is a long way from your kitchen.
You’ve probably seen Optimus by now. It’s white with a black head and torso, and it shows up in clips dancing, passing popcorn, putting trash in a bin, vacuuming and pressing a microwave button. Other clips show it falling backward while handing out water bottles and struggling to iron a shirt.
The sales pitch is enormous
Musk believes Optimus will be “not just Tesla’s biggest product ever, but probably the biggest product ever.” He wants it working on factory floors first and in our homes later. He told shareholders in July that it eventually “will have human and then superhuman dexterity,” and he has argued the robots could automate almost all human labor, from hauling sheet metal to folding laundry, for as little as $20,000 each. He made the end-of-2027 prediction at the World Economic Forum’s annual meeting in Davos, Switzerland, in January.
And he has company. Marc Andreessen, cofounder and general partner of the venture capital firm Andreessen Horowitz, has said robotics could become the “biggest industry in the history of the planet.” In January, Nvidia CEO Jensen Huang said humanoid robots would match human-level ability this year. Morgan Stanley said the number of robots that “resemble and act like humans” is likely to reach nearly 1 billion by 2050, creating a market worth over $5 trillion.
The logic behind all of this goes like this: The AI that powers OpenAI’s ChatGPT and Anthropic’s Claude learned to imitate human language, so a new generation of robots should be able to learn to imitate human movement the same way.
A lot of robotics researchers don’t believe it. Yann LeCun, often called one of the godfathers of AI, said at another Davos event in January that none of the companies building humanoid robots has any idea how to make them smart enough to be useful.
Jonathan Hurst, cofounder and chief robot officer of Agility Robotics and a robotics professor at Oregon State University, said people mix up two separate things. “It’s very easy to make a robot that looks like a person,” Hurst said. “It is dramatically more difficult to make a machine that moves or behaves dynamically or physically like a person.”
That distinction matters. A humanoid robot only has to look like a person. A generalist robot has to learn and carry out lots of different tasks. The hype treats those as the same thing, and in labs around the country the arguments over timelines are drowning out progress that’s slow but real.
Two arms and a lunchbox
If you want to see one of the smartest robot brains working today, skip the humanoids. Look at ALOHA 2, short for “A Low-cost Open-source Hardware System for Bimanual Teleoperation.”
There isn’t much to it: a pair of arms, some grippers and a couple of cameras on a bench top. It sits on one side of an old argument among roboticists. Supporters of humanlike robots say that shape will help machines fit into the world as it’s already built. Critics say it isn’t worth the trouble. ALOHA 2 is the critics’ answer, and Google DeepMind researchers use it in their labs to test Gemini Robotics, their most advanced AI system for robots.
With Gemini Robotics in control, ALOHA 2 starts acting more like a generalist, meaning it can handle any number of tasks it has seen examples of in training. In a video of a lunch-packing exercise, it uses two pincer grippers to gently place a slice of white bread into a Ziploc bag and close it. Then it puts a bunch of grapes in a Tupperware container, secures the lid, carefully moves everything into a lunchbox and zips it shut.
Nobody would call that a good lunch. Still, a robot packing it on its own is a clear step past what was possible even three years ago.
Most of that progress comes from what AI has done to robot policies. A policy is the part that decides how a general-purpose robot reads its surroundings, plans its movements and then carries out the task correctly.
From hard-coded rules to models that watch
Engineers used to write these policies by hand. That meant thousands of lines of code spelling out every millimeter of a robot’s movement across hundreds of tasks. Over the last few years, the work has shifted to advanced AI systems, and that shift explains most of the current optimism about generalist robots.
Vision-language models, or VLMs, came first. They work like large language models but learn from images as well as words. Show a VLM a picture of a coffee spill, ask it to find something to clean the mess, and it can pick out a nearby cloth. When policies were hard-coded a couple of years ago, robots had no grasp of context like that.
Vision-language-action models, or VLAs, add motion commands. A VLA trains on images or videos of a task along with data on how a robot arm moves to do it. That movement data usually comes from teleoperation, where a person uses remote controls to walk a robot through the motion. Put a VLA-powered robot in front of a desk, tell it to “close a laptop” or “wrap up the headphone wire,” and it will scan the scene, find the right object, plan the move and swing its arms into action. That works as long as it has seen the task done before.
Gemini Robotics is a VLA trained on many hours of human demonstrations covering a wide range of actions. That’s why it can pick up snow peas with kitchen tongs, fold origami or pack a simple lunch.
The limitation is a big one. Ask a VLA-controlled robot to do something outside its training set and it’s highly likely to fail.
“Thinking about the space of all tasks, a real generalist policy would be able to do everything along that spectrum,” said Edward Johns, a robotics professor at Imperial College London. A Gemini Robotics model can currently manage only “a few things here and a few things there,” Johns said.
The data problem nobody has solved
The industry’s standard answer is predictable. Give the models more data, which means more examples, which should lead to more generality. Pannag Sanketi, a former tech lead in robotics at Google DeepMind who now works on his own AI robotics project, said the company wants to gather “as much data as possible.”
Finding it is the hard part. Large language models had oceans of existing text to learn from. Nothing comparable exists for high-quality physical demonstrations.
Every workaround has a cost. Paying large numbers of people to produce teleoperation data is expensive and slow. Training VLAs on videos of people doing things produces poor-quality data. Sending robots into the real world to collect experience runs into the problem that robots aren’t safe or reliable outside labs. Sanketi said a “multi-prong” approach that pulls data from all of these sources is the most likely way forward.
Hurst doesn’t think more data fixes anything. He called the idea “a fundamentally flawed premise.”
His argument is that real-world tasks get complicated fast. Take making coffee. Every kitchen is different, coffee machines work in different ways, cups need different grips, and grounds, hot water and milk each have to be handled differently. Getting to generality with VLAs, Hurst said, would require “complete data coverage of all of the things that [a robot] could ever do.” That amounts to an almost infinite supply of training data.
LeCun was blunter in Davos. “The [AI] approaches that have been successful for language do not work for high-dimensional, continuous, noisy data,” he said, referring to the kind of data robots deal with all the time. “You have to use something else.”
World models are getting the money
The front-runner for that “something else” is the world model. It’s AI trained less on text and more on video, 3D scans and sensor data, and it’s built to predict what happens when something acts in the real world. The goal is an internal picture of reality accurate enough to capture how objects move, collide, fall and deform.
That would help in two ways. If simulations matched real physics closely enough, robots could train in them, making development faster, cheaper and safer and cutting down on real-world testing. A robot with a world model could also reason about its surroundings instead of just reacting to them, and predict what an action will do before it takes it.
Nvidia and Google are working on the technology, and investors are pouring money into high-profile startups. World Labs, cofounded by Stanford AI researcher Fei-Fei Li, raised $1 billion in February, and AMD acquired it at the end of September for $8.2 billion. AMI Labs, cofounded by LeCun, formerly Meta’s chief AI scientist, also raised $1 billion in March.
Even the people at the center of it say it’s early. Late last year, Li described the field as “nascent” and said “foundational approaches are still being established.” In a June Substack post, she laid out serious challenges. World models are a promising area of research for now. They aren’t a shortcut to general-purpose robots, though some early results are starting to show what they could do.
A robot almost air-fries a sweet potato
The most interesting of those results came in April from a lab in San Francisco’s Mission District. The startup Physical Intelligence, which goes by PI (as in π), wants to build a universal brain that could, in theory, turn any robot into a generalist. Its approach is to train on everything it can get.
In 2024, PI published details of π0, its first generalist robotics system. It was a VLA the company called the “most capable and dexterous generalist robot policy to date.” π0 first trained on a proprietary set of 10,000 hours of teleoperated human demonstrations plus several open-source robot datasets. π0.5 arrived in spring 2025 and trained on a broader mix of data, including labeled images from the web, which made it more versatile. A fall 2025 update, π0.6, added reinforcement learning.
Each version got better. The model went from slowly folding laundry to putting things away in new environments to folding boxes with a higher success rate.
Then came π0.7 in April 2026. It uses a less powerful world model that generates images of the steps needed for a task. While the robot works, this “lightweight” model feeds it snapshots of what to do next.
PI claims the model shows the first signs of compositional generalization. That’s when an AI system performs a skill it has never been exposed to by recombining skills from its training data. In one test, PI asked the model to “load a sweet potato into the air fryer,” a task it had never seen. In the demo video, the robot fumbles a bit, makes a few false starts and eventually puts in a reasonable effort. It doesn’t fully finish.
Sergey Levine, a professor at the University of California, Berkeley, and a PI cofounder, is enthusiastic about it. “It’s actually the first time that we’ve convincingly seen that kind of compositional generalization, where we can basically ask the model to do tasks that we did not specifically collect data for and train it to do, and it’ll actually make a passable attempt,” Levine said.
The follow-up is the part I find most revealing. The team dug through the training material to figure out how the model managed it. They found bits of relevant labeled teleoperation data, including two examples of a human operator using the robot to push an air fryer basket into the fryer. Those two scraps may have been enough to get π0.7 most of the way to cooking a sweet potato.
It’s still unclear how impressive π0.7’s ability to generalize really is. But since the model saw air fryers only briefly, the result shows roughly how far the most advanced research can take robots today.
The man behind the curtain
You may notice a gap between these small lab wins (“Look! It put a sweet potato into an air fryer!”) and the lifelike agility in viral demos, where robots dance on stage and politely serve drinks. Those demos often leave out a key fact. In many cases, a human is controlling the robot or has carefully scripted what it does.
The robot that appeared onstage with Huang in March 2025, seemingly following his instructions and trailing him around, was remote-controlled by what its makers called “a puppeteer behind the scenes.”
Fully autonomous motion planning, where a robot works out on its own where to go, is still a largely unsolved problem, especially in new and chaotic places like a construction site or an unfamiliar home. A bigger and often related problem is getting robots to take on larger, vaguer jobs made up of several tasks, where the robot has to decide how to do each one and in what order. A robot that can put a plate in the microwave is a long way from one that hears “Make dinner,” checks the refrigerator, chops ingredients and turns on the stove. The risotto test mentioned at the top is Google DeepMind’s attempt at that kind of job, and so far it has failed.
Reliability is another problem. A useful robot has to get it right essentially every time. Marc Raibert, founder of Boston Dynamics, said VLA researchers celebrate the wrong milestones. “People are very excited when their result goes from 50% success to 70% success,” Raibert said. “But 70% success is like it doesn’t work, right?”
What’s working in the field is boring
The few humanoids being tested in real-world settings do very limited jobs in tightly controlled environments. None of them are generalists.
Agility said it has hundreds of robots deployed in trials at facilities owned by GXO Logistics, Amazon and Schaeffler. For now, Hurst said, they handle simple tasks like moving bins and totes. And it took years to build robots safe enough for logistics companies to even consider using them, he said.
Musk’s track record here is worth a look. In May 2025, he claimed “thousands” of Optimus robots would be working in Tesla factories by the end of that year. In January of this year, he said the company had only “some of the Tesla Optimus robots doing simple tasks in the factory.”
The $20,000 robot in your living room still needs a human
Factories are hard. Homes are harder. You can preorder the 1X Neo home robot right now, and it’s expected to ship sometime later this year for $20,000. It promises to handle “the boring and mundane tasks around the house,” like putting away dishes, answering the door and tidying the living room, “so you can focus on what matters to you.”
The plan is for the five-foot-six-inch robot to eventually do all of that by itself. For now, it needs a remote human operator for most tasks. That means you’d have to give a stranger permission to look around your home through the robot’s cameras.
When asked how long it will take for fully autonomous robots to be ready for household work, Hurst gave a number. “If I had to pick a number, I’d say it’s 10 years before robots are … actually doing useful things in people’s homes,” Hurst said.
When that day arrives, the robots may well be Chinese. Nearly 90% of the roughly 15,000 humanoid robots shipped in 2025 came from Chinese companies, according to market intelligence firm Omdia and Chinese robotics company Unitree. Unitree shipped more humanoids than any other company last year, and one of its models costs less than $6,000. That undercuts Musk’s best-case Optimus price by more than $14,000, though Unitree expects its machines to go into industrial work first. The AP recently reported that the buyers are mostly corporate and academic labs and state-owned enterprises.
We’ve seen this demo before
People have chased humanoid robots for a long time. In 1495, Leonardo da Vinci sketched a mechanical knight run by cables and pulleys. Westinghouse’s Elektro, a seven-foot-tall box on legs, smoked a cigarette at the 1939 New York World’s Fair. WABOT-1, built by Japan’s Waseda University in 1973, was the first full-scale programmable humanoid.
Honda’s ASIMO, unveiled in 2000, was probably the first one that was at all competent. To some degree, it could climb stairs, recognize faces and move through spaces on its own. Honda discontinued it in 2018 because it never got far enough past its demos to be useful.
Every one of those machines impressed people in its day, and none of them could handle the real world. The air fryer test is the same story. Today’s robots, with all the AI behind them, still get stuck on the step that sends them from a staged demo into a messy kitchen they’ve never seen.
