'70% success is like it doesn't work': why AI robots aren't coming to your home

Elon Musk told shareholders in July that Tesla's Optimus humanoid robot "will have human and then superhuman dexterity," and he predicted at Davos in January that it could be on sale to the public by the end of 2027. Yet in January of this year, Musk admitted that Tesla had only "some of the Tesla Optimus robots doing simple tasks in the factory" — a long way from the "thousands" he claimed in May 2025 would be working at Tesla factories by the end of that year.

That gap between projection and reality runs through the entire humanoid robotics industry, as detailed in a recent MIT Technology Review investigation. The stakes are enormous: Morgan Stanley projects the number of robots that "resemble and act like humans" will reach nearly 1 billion by 2050, creating a market worth over $5 trillion.

What are industry leaders promising?

The optimists are lining up. Musk calls Optimus "not just Tesla's biggest product ever, but probably the biggest product ever," arguing the robots could automate almost all human labor for as little as $20,000 each. Marc Andreessen, cofounder of Andreessen Horowitz, has said robotics could become the "biggest industry in the history of the planet." Nvidia CEO Jensen Huang said in January that humanoid robots would match human-level ability this year.

The skepticism is equally blunt. "None of those companies [building humanoid robots]—absolutely none of them—has any idea how to make those robots smart enough to be useful," Yann LeCun, often described as one of the godfathers of AI, said at a Davos event in January. "The [AI] approaches that have been successful for language do not work for high-dimensional, continuous, noisy data. You have to use something else."

Researchers also warn against conflating human-looking machines with generalist ones. "It's very easy to make a robot that looks like a person," explains Jonathan Hurst, cofounder and chief robot officer of Agility Robotics and a robotics professor at Oregon State University. "It is dramatically more difficult to make a machine that moves or behaves dynamically or physically like a person."

How smart are robot brains today?

The cutting edge is visible at Google DeepMind, where researchers use a device called ALOHA 2 — a pair of arms, grippers, and cameras — to test their Gemini Robotics system. Controlled by Gemini Robotics, ALOHA 2 can pack a lunchbox: place bread into a Ziploc bag, seal it, put grapes in a Tupperware container, secure the lid, and zip everything into the lunchbox.

It's not a great lunch. But it represents an objective step forward from what was possible three years ago.

The progress stems from a shift in how robots are controlled:

• Hard-coded policies (old approach): Engineers wrote thousands of lines of software dictating each millimeter of a robot's movements.
• Vision-language models (VLMs): Trained on images and words, these let a robot look at a coffee spill and identify a nearby cloth to clean it.
• Vision-language-action models (VLAs): These add motion commands, trained on video of tasks plus teleoperation data showing how a robot arm should move to perform them.

Gemini Robotics is a VLA, trained on many hours of human demonstrations. It can pick up snow peas with kitchen tongs, do origami, and assemble a simple lunch. The limitation is glaring: ask a VLA-controlled robot to perform a task outside its training set, and it will most likely fail.

"Thinking about the space of all tasks, a real generalist policy would be able to do everything along that spectrum," says Edward Johns, a robotics professor at Imperial College London. Today, Gemini Robotics can do only "a few things here and a few things there."

Is more training data the answer?

The standard response is to train on more data. But large language models had oceans of existing text; no comparable pool of high-quality physical demonstrations exists. Researchers have three imperfect options:

• Employing large numbers of people to collect teleoperation data — costly and time-consuming.
• Training on videos of people performing activities — the resulting data quality is poor.
• Deploying robots in the real world to learn from experience — robots aren't yet safe or reliable outside labs.

Pannag Sanketi, a former tech lead in robotics at Google DeepMind, thinks a "multi-prong" approach drawing on all these sources is the most likely path forward.

Others reject the data-scaling premise outright. Hurst calls it "a fundamentally flawed premise," arguing that achieving generality through VLAs would require "complete data coverage of all of the things that [a robot] could ever do" — an almost infinite pool of training data, because real-world tasks explode in complexity: no two kitchens are identical, coffee machines work differently, and cups require different grips.

Why are world models attracting billions?

The leading alternative is the world model — AI trained on video, 3D scans, and sensor data, built to predict the outcomes of actions in physical reality. Nvidia and Google are both developing the technology, and investors are betting heavily:

• World Labs, cofounded by Stanford's Fei-Fei Li, raised $1 billion in February and was acquired by AMD at the end of September for $8.2 billion.
• AMI Labs, cofounded by LeCun, raised $1 billion in March.

Even the field's champions urge caution. Li described the discipline as "nascent" late last year, adding that "foundational approaches are still being established."

Did Physical Intelligence just show a breakthrough?

In April 2026, the San Francisco startup Physical Intelligence released π0.7, the latest version of a generalist robotics system first published as π0 in 2024, trained on a 10,000-hour proprietary collection of human demonstrations. The spring 2025 update, π0.5, added labeled images from the web; the fall 2025 π0.6 added reinforcement learning.

π0.7 incorporates a lightweight world model that generates images of the steps needed to perform a task, feeding the robot snapshots of what to do next. The company claims the model shows the first signs of compositional generalization — performing skills it was never explicitly trained on by recombining learned ones.

In one test, researchers asked the model to "load a sweet potato into the air fryer," a task it had never encountered. The robot fumbled, made false starts, and eventually produced a reasonable, if incomplete, attempt. "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," said Sergey Levine, a UC Berkeley professor and PI cofounder.

The team later found snippets of relevant teleoperation data in the training material — including two examples of a human pushing an air fryer basket into the fryer — which may explain the near-success. Just how general π0.7's abilities are remains unclear.

How much of what we see is real?

Many viral robot demonstrations obscure a key fact: humans often control the robot or carefully script its actions. The robot that appeared onstage with Nvidia's Jensen Huang in March 2025, seemingly responding to his instructions, was remote-controlled by what its makers called "a puppeteer behind the scenes."

Reliability is the deeper problem. "With VLAs, people are very excited when their result goes from 50% success to 70% success," says Marc Raibert, founder of Boston Dynamics. "But 70% success is like it doesn't work, right?"

Deployed humanoids remain limited. Agility has hundreds of robots in trials at facilities owned by GXO Logistics, Amazon, and Schaeffler, but Hurst says they perform simple tasks like moving bins and totes — and it took years to make them safe enough for logistics firms to consider using them at all. Google DeepMind's attempt to have a robot survey a kitchen and pack ingredients for mushroom risotto into a basket ended in failure.

When will robots reach homes?

The 1X Neo home robot is available for preorder at $20,000, with delivery expected later this year. It promises to handle "the boring and mundane tasks around the house," but for now a remote human operator is needed for it to do most things. "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.

China dominates production: nearly 90% of the roughly 15,000 humanoid robots shipped in 2025 were made by Chinese companies, according to Omdia and Unitree. Unitree, which shipped more humanoids than any other company last year, sells a model for under $6,000 — though it expects its machines to be used in industrial applications first, and the AP reported that orders come predominantly from corporate and academic labs and state-owned enterprises.

The pattern is an old one. Leonardo da Vinci sketched a mechanical knight in 1495. Westinghouse's Elektro smoked a cigarette at the 1939 World's Fair. Waseda University built the first programmable humanoid, WABOT-1, in 1973, and Honda's ASIMO, discontinued in 2018, could climb steps and recognize faces but never advanced far beyond demonstrations. Each was a feat of engineering; none could navigate the real world. Whether today's AI-powered machines break that pattern may depend on whether world models — not bigger datasets — can finally give robots physical intuition.