Aiodic

Back to the table

114 Emerging Eb Embodied AI Frontier

Embodied AI

AI in robots and physical bodies.

What it is

Embodied AI is artificial intelligence that has a body. Instead of living only on a screen, answering questions or writing text, it sits inside a machine that can sense the world and move through it: a robot arm in a warehouse, a self-driving car, a vacuum that maps your living room, or a humanoid robot with two arms and two legs.

Every embodied system has to do three things, over and over, many times a second:

A chatbot only does the middle step. If it gets something wrong, you read a bad sentence and ask again. An embodied system can get any of the three wrong in real time, in a world that doesn't wait. The cup slides, the child steps off the kerb, the shirt is inside out. That's why embodied AI is harder than it looks, and why it sits in the frontier section of this table.

Moravec's paradox: the hard things are easy, and the easy things are hard

In the 1980s, the roboticist Hans Moravec pointed out something that still surprises people. In his 1988 book Mind Children, he observed that it was comparatively easy to get computers to perform like adults on intelligence tests or at checkers, and difficult or impossible to give them the perception and mobility skills of a one-year-old. The idea became known as Moravec's paradox.

You can see it in your own day. A chatbot can pass hard exams, write a legal summary, and explain quantum physics. But folding a basket of laundry, which a ten-year-old can do while chatting, is still a research challenge for robots. Think about what folding actually involves: finding the corner of a crumpled towel you can barely see, feeling how much it stretches, adjusting your grip as it slips, and recognising the moment a shirt is the right way out. None of that was ever written down in a way a computer could read. Humans learned it through years of handling things, and evolution spent millions of years building the brain machinery for it.

The usual explanation is that the skills we find hard, such as algebra or chess, are recent and narrow, and we do them consciously, so we can describe the steps. The skills we find easy, such as seeing, grasping, and balancing, are ancient and happen below awareness, so we can't describe them, and neither could the early programmers who tried to write them as rules.

How modern embodied AI works

For most of robotics history, engineers programmed robots step by step: move here, close the gripper, lift ten centimetres. That works well in a factory where every part arrives in the same place, and poorly anywhere messy. The recent shift borrows the recipe that made language models so capable: train one large model on lots of varied data and hope general skills emerge.

Why it matters

Most of the economy is physical. Language models have changed office work, but most jobs involve moving things: stocking shelves, cleaning rooms, cooking, caring for people, building houses, delivering parcels. If AI is going to make a difference to those jobs, it needs a body. That's why so much investment is now flowing into what the industry has started calling "physical AI".

It's where the hardest open problem is: data. Language models learned from a large share of the text humans have ever put online. There is no internet of robot actions. Nobody has been recording, for decades, what every joint of every hand did while tying shoelaces. The biggest shared robot dataset, Open X-Embodiment, released in October 2023, pooled more than a million recorded robot episodes from 22 kinds of robot and 21 institutions. That was a landmark for robotics and is still tiny next to the trillions of words behind a chatbot. Almost every strategy in the field is an answer to this shortage: collect demonstrations by teleoperation, generate practice in simulation or with synthetic data, learn from human videos, or put robots into real homes and factories so that they collect data as they work.

Mistakes can hurt people. A chatbot that hallucinates gives you a wrong fact. A robot that misjudges gives someone a bruise, or worse. Industrial robots have traditionally worked in cages or behind light curtains that stop them when a person steps in. Robots meant to work beside people, in homes and on shop floors, have to be slow enough, soft enough, or clever enough to stay safe without the cage. That makes safety engineering, testing, and guardrails much more important than in software alone, and it's why regulators treat self-driving cars and robots differently from apps.

It separates real progress from theatre. Robot videos are easy to stage. A clip can be sped up, cut from the one success out of twenty tries, or secretly driven by a human. Knowing how embodied AI works helps you ask the right questions: is this autonomous or teleoperated, is it one take or many, and is it a product or a demo?

It's where several frontier ideas meet. Embodied AI pulls together planning (deciding a sequence of steps), world models (predicting what happens next), continual learning (improving from experience after deployment), and autonomy levels (how much a person stays in control). Some researchers argue that a body is needed for anything like AGI, because so much human understanding comes from physical experience. Others disagree. It's an open question, and worth treating as one.

Real-world examples

It helps to sort embodied AI into three groups: what works at scale today, what has been shown in research, and what is still mainly demos and promises.

Working at scale: robotaxis and warehouses (2024 to 2026)

The clearest success of embodied AI so far is on the road. Waymo, the self-driving company owned by Google's parent Alphabet, runs taxis with no driver in the seat. In May 2024 it was giving about 50,000 paid rides a week. By March 2026, Waymo said it was giving 500,000 paid rides a week across ten US cities, as reported by TechCrunch. In February 2026, co-chief executive Tekedra Mawakana told Bloomberg the company was aiming for one million paid rides a week by the end of 2026. That's a goal, not a result, so check the latest figures yourself.

It hasn't been flawless. In December 2025 Waymo issued a software recall covering 3,067 vehicles after its cars were found to pass stopped school buses, including incidents documented by the school district in Austin, Texas. The fix was a software update, which is itself a lesson about embodied AI: every car learns the correction at once, but every car also shared the mistake.

Warehouses are the other place robots already work in huge numbers, largely because the environment can be designed around them. In July 2025 Amazon said it had deployed its millionth robot, across more than 300 facilities worldwide, and introduced an AI model called DeepFleet that coordinates how its robots move around a building, which Amazon said would improve travel time across the fleet by 10 percent. Most of those robots are not humanoids. They're shelf-lifting drive units, sorting arms, and carts, doing narrow jobs in spaces built for them.

Notice the pattern. The successes are in places where the task is well defined and the world can be mapped, fenced, or standardised. Roads have rules and lane markings. Warehouses have barcodes and fixed layouts.

In research: robots that follow instructions and fold laundry (2023 to 2025)

On 28 July 2023, Google DeepMind introduced RT-2, the model that gave vision-language-action models their name. The trick was to write robot movements as if they were words, so a model trained on web images and text could learn to "say" actions the same way it says sentences. Across about 6,000 trials, DeepMind reported that performance in situations the robot had never seen rose from 32 percent with its previous robot model to 62 percent. In one demonstration, asked to pick up an object that could be used as an improvised hammer, it picked up a rock. That's the internet's knowledge showing up in a robot's hand.

On 31 October 2024, the start-up Physical Intelligence published π0 (pronounced "pi-zero"), a general-purpose robot model trained on over 10,000 hours of robot data from seven different robot setups, plus open datasets. Its most striking demonstration was laundry: taking crumpled clothes out of a bin, or a dryer, and folding them. For a Moravec's paradox task, that was a real step. The researchers were also candid in their paper that not all tasks in their evaluation worked reliably, and that it wasn't yet clear how much data would be needed to reach near-perfect performance.

In March 2025, Google DeepMind followed with Gemini Robotics, a VLA built on its Gemini 2.0 model, demonstrated on tasks such as folding origami and packing a snack into a zip-lock bag. It also released a dataset for testing whether robots understand which actions are unsafe.

These are real research results, published with methods and numbers. They are also research: controlled settings, chosen tasks, and success rates that are impressive for robots but not yet good enough for a product you'd leave alone with your shirts.

Mostly demos and promises: humanoids in factories and homes (2024 to 2026)

Humanoid robots, with a human-like body, attract the most attention and the most disputes.

On 10 October 2024, Tesla held its "We, Robot" event, where Optimus humanoid robots walked among guests, served drinks, and chatted. Bloomberg, The Verge, and others then reported that the robots were being controlled in part by humans working remotely, and one robot told a guest on video that it was "assisted by a human". Reports said the walking was done by the robots' own AI, while the conversations and much of the handling relied on people. Tesla had not said this at the event.

Figure AI, a heavily funded humanoid start-up, announced a deal with BMW in 2024 to test its robots at BMW's plant in Spartanburg, South Carolina. In April 2025, Fortune reported that BMW described a more modest picture than Figure's chief executive Brett Adcock had given publicly: a single robot at a time, at first practising during non-production hours. Adcock rejected the report and threatened legal action. In November 2025, Figure said its Figure 02 robots had completed an 11-month deployment at the plant, running 10-hour shifts, loading more than 90,000 sheet-metal parts, and contributing to the production of more than 30,000 cars. Those figures come from Figure itself. Days later, a former Figure safety engineer sued the company, alleging he was fired after warning that its robots were powerful enough to cause serious injury to a person; Figure denied wrongdoing and said he was dismissed for poor performance. The case shows how hard it is, from the outside, to tell where a humanoid programme really stands.

Homes are further off still. At the end of October 2025, the Norwegian-founded company 1X opened pre-orders for NEO, a humanoid meant for household chores, at $20,000 or $499 a month at the time of the announcement. 1X was open that NEO would rely on remote human operators, which it calls Expert Mode: for chores the robot can't yet do alone, a 1X employee can take control through a virtual reality headset, with the owner's permission, and the recordings help train the robot. 1X said first deliveries would come in 2026. In May 2026 it opened a factory in California and said early robots were going first to internal testing. Check the current state before assuming these robots are in homes, and read the privacy terms carefully: a stranger seeing through a camera inside your home is a real trade-off.

A scenario: the robot that could explain folding but couldn't fold

This is an illustrative scenario, not a real event. Imagine a family trying a home robot demo. They ask the robot's voice assistant how to fold a fitted sheet, and it gives a clear, correct, five-step explanation. Then they ask it to actually fold one. It picks up a corner, loses it, picks up the middle by mistake, and after four minutes asks whether they'd like to connect a remote operator. That gap, between knowing about the world and handling it, is Moravec's paradox in a single afternoon.

Try it yourself

You don't need a robot for this. The goal is to build your own feel for Moravec's paradox and to practise asking "is that really autonomous?"

  1. List ten tasks from your day. Write down ten things you did today, mixing physical and mental: making coffee, replying to an email, tying shoelaces, planning a route, washing up, writing a shopping list, putting away groceries, reading a bus timetable, folding a towel, deciding what to cook.
  2. Rate each task twice. Give each task two scores from 1 (easy) to 5 (very hard): one for how hard it would be for a chatbot, and one for how hard it would be for a home robot to do reliably in your home. Do this before asking any AI, so your own guesses are on paper.
  3. Ask a chatbot to rate the same list. Paste this into any chatbot, replacing the list with yours: > Here are ten everyday tasks: [paste your list]. For each one, rate from 1 (easy) to 5 (very hard) how hard it is today for (a) a text chatbot and (b) a general-purpose home robot to do reliably in an ordinary home. Give a one-sentence reason for each rating. Then tell me which tasks are the clearest examples of Moravec's paradox, and why.
  4. Compare. Where did you and the chatbot disagree? People often underrate how hard physical tasks are, because they feel effortless to us. That surprise is the whole point of the exercise.
  5. Check one claim against a source. Pick the physical task the chatbot rated easiest for a robot, and search for real evidence: a research paper, a company announcement, or a news report. Ask yourself whether it shows a product, a lab result, or a demo video. The chatbot can help you plan the search: > I want to check whether robots can really [task] reliably today. Suggest three search queries I could use, and tell me which kinds of sources would count as strong evidence (for example a published paper with success rates) versus weak evidence (for example a promotional video).
  6. Run the demo checklist on a robot video. Next time you see an impressive robot clip online, ask these five questions, or paste them into a chatbot along with the video's caption: > Help me judge this robot video. Based on the caption and any details given: 1) Does it say the robot was fully autonomous, teleoperated, or doesn't it say? 2) Is the video sped up? 3) Is it one continuous take or edited? 4) Is this a product customers can buy or a research demo? 5) What evidence would I need to believe the robot can do this reliably? Caption: [paste caption] Remember that the chatbot can only work from what you give it, and it can be wrong. Treat its answer as a list of questions, not a verdict.

Common mistakes and misunderstandings

Assuming a smart chatbot means a smart robot is close. Knowing how to fold a shirt and being able to fold one are different skills. Language models solved the "knowing" part surprisingly fast; the "doing" part moves more slowly because the data and the hardware are harder.

Treating demos as deployments. A video of a humanoid unloading a dishwasher shows that it happened at least once. It doesn't tell you how often it fails, how long it took, or whether a person was helping. Look for success rates, continuous footage, and independent reporting.

Missing the human in the loop. Many robots, from home humanoids to some delivery robots, rely on remote operators for part of the job. That is often sensible and honest, and it's how the robots collect training data. It becomes a problem only when it's hidden. Ask whether a human in the loop is involved, and how much.

Thinking a humanoid shape is the point. The robots doing the most useful work today, in warehouses and on roads, look nothing like people. The argument for humanoids is that our world is built for human bodies, so a human-shaped robot could use our stairs, tools, and kitchens. It's a reasonable bet, not a proven one.

Believing forecasts because of who makes them. At CES in January 2025, NVIDIA's chief executive Jensen Huang said that "the ChatGPT moment for general robotics is just around the corner." Many company leaders have made similar predictions about home robots and driverless cars, some of them many years ago. Predictions from people selling chips, cars, or robots are worth hearing and worth checking. Look at what has shipped, not what has been promised.

Forgetting that software updates change physical behaviour. When a robot fleet gets a new model, every machine can change how it acts overnight, for better or worse, as Waymo's school bus recall showed. Testing and gradual rollouts matter even more than in apps.

Overlooking privacy. A robot in your home has cameras and microphones that move around, and some send data back to the company or let remote staff see through them. Read the privacy terms before you buy, and check them again after updates, because terms change.

Thinking it's all or nothing. Embodied AI is arriving in pieces: a car that drives itself in some cities, a vacuum that avoids socks, an arm that picks parcels. Waiting for a robot butler misses the real, narrower changes already happening.

Related elements