The remote-controlled economy arrives, just in time to end

TL;DR

• Robots that a person drives from far away are now a real business. Tau rents humanoid cleaners in Los Angeles for $30 an hour, and a person steers each one from somewhere else.
• It only pays when the robot does most of the work. Telexistence's arms restock drinks fridges in Japan on their own 96-98% of the time, so one pilot in Manila watches about 50 of them.
• The best measure of autonomy is how many robots one pilot runs. One pilot for 10 robots means each robot works alone at least 90% of the time.
• Today's remote-driven humanoids are too slow to beat a worker, even with no pilot. At 30% of human speed (our estimate), a $50,000 humanoid costs $31 for one hour of a person's work, more than a Danish night stocker earns ($27.90).
• Our view: by late 2027, robots do basic chores with no pilot. The pilot then becomes a rescue desk, like Waymo's remote agents today.

In an office building in Manila, about 60 people sit in front of screens and VR headsets. Each pilot watches about 50 robot arms that restock drinks fridges in Japanese convenience stores, 3,000 km away. When an arm gets stuck (a bottle falls and rolls away, for example), the pilot takes control, fixes the problem (it can take up to five minutes) and lets go. In a typical eight-hour shift, each pilot does this about 50 times. The pilots earn $250 to $315 a month, "roughly the same as a call center agent" (Rest of World, last October), and one of them told the reporter they often feel "dizzy and cross-eyed" in the headset.

In 2026, remote-driven robots went far beyond fridges, into homes, supermarkets, factories and apartment blocks. Tau Robotics rents humanoid cleaners in San Francisco and Los Angeles for $30 an hour, and its terms are very plain: "The robot is not autonomous. A person is controlling it." The Danish supermarket chain føtex books shifts from HIVE Robots, whose humanoids an operator steers with a VR headset. Undaunted rents security robot dogs to apartment complexes in Atlanta (a remote operator watches through them, and talks through them). And Nucleus rents humanoids to a German factory by the hour.

Many people in robotics see the remote pilot as a bridge: people drive the robots today, and the robots learn from them to work alone tomorrow. Is it? Yes, but only for a short part of the way. The pilot is what lets a customer rent a robot today (a person guarantees the job gets done), and the pilot rescues robots when they're stuck. But the robots mostly learn somewhere else (from the AI labs' models, from videos of people, and from their own practice), and a pilot business only makes money once the robot works mostly alone.

Forty years of avatars

Marvin Minsky, one of the founders of AI, described the Manila pilots' world in June 1980, in Omni magazine: "We could have a remote-controlled economy by the twenty-first century if we start planning right now." He added: "Manual labor could easily be done without leaving your home." He called it telepresence, and gave the credit for the dream to Robert Heinlein's 1942 story "Waldo" (its hero works through remote-controlled hands, and engineers still call those "waldos").

The same year as Minsky's essay, in Japan, the engineer Susumu Tachi gave remote-controlled robots their own name, "telexistence", and he spent his career on it. For the next 40 years, remote robots made money only where the job needs an expert's hands, or where you have to keep a person out of danger. That means surgery (Intuitive's da Vinci, since 2000), bomb disposal and subsea oil. For ordinary work, the robots worked technically, but the businesses failed:

• Toyota's T-HR3 (2017) copied an operator's whole-body moves, and in 2018 someone drove it over 5G from 10 km away. It stayed a demo.
• The ANA Avatar XPRIZE (a contest for the best robot avatar) gave out $10 million in 2022, and NimbRo (from the University of Bonn) won the $5 million grand prize. No avatar business of note came out of it.
• Sanctuary AI put a humanoid in a Canadian store for a week in 2023, driven by a "human minder" in a pilot rig. It reported "110 retail-related tasks completed correctly", and Sanctuary sold it as "labor as a service". Its founding CEO was removed in 2024. Since June, Sanctuary sells its AI to run on existing factory robots (only a proof of concept so far).
• Phantom Auto drove vehicles and forklifts remotely. It raised about $95 million, and shut down in March 2024.
• Amazon's Just Walk Out stores (the shops with no checkout) used about 1,000 reviewers in India. In 2022 they reportedly checked 700 of every 1,000 sales, when the goal was 50 (Amazon disputes this).

The lesson from all of them: when the person does most of the work from far away, you pay for the person and for the machine. It's a call centre, with a very expensive robot attached.

The economist Richard Baldwin saw what changes once the robots get good. In The Great Convergence (2016), he predicted that remote-controlled robots would let people do physical work in another country. In 2019 he gave a name to the people who already work in another country's offices without leaving their own: "telemigrants". Their only limit, he said that year: "There's nothing they can't do except be in the room." Remote-controlled robots remove that limit.

An arm behind a fridge

Telexistence is the company that found a way that works: let the robot do the job, and keep the person for the rare problems. It started in Tokyo in 2017, with Tachi as its chairman. Its first robot, Model H (2018), was a full avatar that you could "see, hear, and feel through". In 2020 it tried Model-T, a remote robot with 22 joints, in FamilyMart and Lawson stores, driven from its office about 8 km away. It didn't scale.

So in 2022 it shipped something much simpler: TX SCARASelective Compliance Assembly Robot Arm: a fast arm that moves flat, in a horizontal plane. Common in factories for picking and placing., an arm on a rail behind the drinks cooler, which does only one job (restocking bottles and cans). This arm works alone 96-98% of the time. At that level one pilot can run about 50 robots, and that's when a pilot office in Manila started to make sense. In 2025 Telexistence partnered with Physical Intelligence (a startup that builds AI models for robots), to learn from its pilots' corrections and become fully autonomous. It's also building a humanoid for Seven-Eleven Japan, planned for 2029. ## Robots per pilot

Robotaxis show the same thing, with more data. Pony.ai went from one remote assistant for every 3 cars to one for every 20, in a single year. WeRide went from 1:10 to 1:40. And in February, Waymo ran 3,000 cars and over 400,000 rides a week with about 70 remote agents on duty (half of them in the Philippines).

These agents don't drive the cars. They "provide guidance, they do not remotely drive the vehicles", Waymo's safety chief, Mauricio Peña, said in February. Cruise's CEO said in 2023 that its cars needed remote help 2-4% of the time (so one person could watch about 25 to 50 of them). And these ratios came after a decade of work on autonomy. Google's project started in 2009, and in 2021 Waymo's CEO, John Krafcik, still called it "an extraordinary grind" (harder, he said, than putting a rocket in orbit).

The maths behind the ratio is simple. If a robot needs a person 10% of the time, one person can run at most 10 robots. If it needs a person 2% of the time, at most 50. So one pilot for 10 robots means the robots work alone at least 90% of the time, and one for 50 means at least 98%. This is why the ratio is such a good measure of autonomy: it tells you how much work a person still does.

And the ratio goes up before full autonomy, because the person's job shrinks step by step, from driving the robot to just pointing at what to do next. A teleoperated Unitree keeps its balance by itself while the operator moves its arms. At Tesla's We, Robot event in 2024, Optimus walked on its own while people remotely ran many of its interactions (Bloomberg reported it). Waymo's agents only answer the car's questions. Each of these steps multiplies how many robots one person can run.

Humanoid services today are at the very start: Tau and HIVE run one pilot per robot, and Nucleus says a two-hour factory demo of its robots was about 60% autonomous. Robotaxis took a decade to climb from there. Humanoids will climb faster, because they don't start from zero: their brains come pretrained by the AI labs, and videos of people and practice in the field teach them the rest.

Failures still come all at once, though. During San Francisco's blackout last December, over 1,500 Waymos asked for help at the same time, and 62 had to be fetched by hand. So the ratio gives you the average, but the worst night tells you how many people you need on staff.

When does a remote-driven robot pay?

A remote-driven robot pays when it's fast enough, or cheap enough. How many robots one pilot runs matters less than you'd think, because remote-driven robots are slow, and a slow robot does less work for the same cost. When the Wall Street Journal tried 1X's NEO last year, a person in VR drove every task, and fetching a water bottle from a fridge 10 feet away took "over a minute". Euronews reported this month that HIVE's robots are still slower than føtex's own staff.

So the fair way to compare is the cost of one hour of work, as a person would do it. A robot at 30% of human speed needs more than three hours to do one hour of a person's work. Here's what that costs for a $50,000 humanoid (what Tau says each of its robots costs to make) and for a $20,000 one like 1X's NEO, both with an operator in the US. The two people they compete with are a Danish night stocker ($27.90 an hour) and a Los Angeles cleaner ($20.93):

At 30% of human speed (roughly where remote-driven robots are today, by our estimate), the $50,000 robot costs $31 for one hour of a person's work, even with no operator at all. So it loses to both workers, whatever the ratio. At 60% speed, it beats the Danish stocker once one pilot runs 6 robots, and the LA cleaner once one pilot runs 14. At full human speed, it only needs 3 robots per pilot to beat the stocker, and 4 to beat the cleaner. And a cheaper body changes a lot: at $20,000, a robot at 30% speed beats the stocker once one pilot runs 10. (All this is before the other costs, like a person on site, battery swaps and supervisors.)

Change the numbers yourself:

With an operator in Manila (about $1.80 an hour), even one pilot per robot pays, once the robot reaches about half of human speed. That's exactly Baldwin's telemigrants: people working in another country without leaving their own, except this time with robot hands. Regulators are already on it: from July 2028, California requires robotaxi remote drivers to be in the US. Tau and 1X already keep their operators in the US, by policy. And latency (the delay between the pilot's move and the robot's) pushes the pilots who drive robots by hand closer to them anyway, into regional hubs.

So it all comes back to the robot. Autonomy removes the cost of the operator, and after that, the robot's speed and the price of its body decide whether it beats a person. At $10,000, a robot at 30% speed costs about $7 per hour of a person's work, with nobody at the controls.

What the pilot teaches

Driving robots by hand had a second promise: data. Every shift a person drives a robot, the robot records how the job is done, and it can learn from that. For a few years, that's mostly how robots learned.

Learning from a person's corrections goes back to DAggerDataset Aggregation (Ross, Gordon and Bagnell, 2011): the learner acts, an expert says what it should have done in the situations it got into, and the learner retrains on all of it. (2011), which showed that a learner improves most when an expert corrects it in the situations it gets itself into (its own mistakes teach it more than the expert's perfect runs). In 2022, a study called "Robot Learning on the Job" (Sirius) had robots keep learning while they worked under human monitors, with extra weight on the moments a person stepped in. Then in 2023 ALOHAA cheap two-arm rig for driving robots by hand (Stanford, 2023): a person moves a pair of "leader" arms, and the robot's arms copy them., a two-arm rig for driving robots by hand, made it cheap ($20,000), and driving robots by hand became the main way to teach them. Physical Intelligence's first model, π0Physical Intelligence's first robot model (October 2024): a vision-language model with 3 billion parameters, plus an "action expert" that sends motor commands up to 50 times a second., trained on about 10,000 hours of robot data in 2024, mostly recorded that way.

In 2025 the biggest bets moved away from it. Tesla mostly switched Optimus training from motion-capture suits and teleopTeleoperation: a person drives the robot from somewhere else, usually with a VR headset or a copy of the robot's arms. Every move can be recorded to train the robot. to video, Business Insider reported, to "scale data collection more quickly". The inventors of ALOHA and of the UMIUniversal Manipulation Interface: a gripper with a camera that people hold in their hand to record how they do tasks, so robots can learn without a robot. gripper (a handheld gripper that records how people do a task) founded Sunday, which trains "solely on data from our Skill Capture Glove", a $200 glove instead of a $20,000 rig. Even 1X, which offers teleop in NEO, tuned its world modelA model that predicts what happens next in a scene, frame by frame, for a given action. Robots use one to imagine the result of a move, or to make training data. on 900 hours of first-person human video and only 70 hours of robot data. Driving by hand still gets the work started: AgiBot, a Chinese humanoid maker, started its study with 336 hours of demonstrations by human experts (the latest version came out in September). And as Sergey Levine (a co-founder of Physical Intelligence) says, "the robot won't master the real world unless it gets to see itself doing stuff in the real world".

What works best is a person stepping in exactly where the robot fails, and then the robot learning from its own runs:

Two recent studies show it. In DITTO-X (Stanford and Columbia), the success rate on a battery task went from 40% to 90% in two rounds, when the person took over from wherever the robot was. (DITTO-X's glove first moves the person's fingers into the robot hand's position, so the takeover is smooth.) With an ordinary hand-tracking glove instead, it got stuck at 47%. And in AgiBot's test on 16 robots, learning from human corrections reached 85%, and adding reinforcement learning on all of the fleet's own runs (takeovers included) took it to 95%. Physical Intelligence wrote that corrections "can only take us so far". And its co-founder Chelsea Finn added this month: "90% success means the person spends >90% of the time passively watching the robot". In other words, once the robot mostly works, the pilot spends most of the shift waiting, which is an expensive way to collect a few corrections.

So driving robots by hand is an expensive way to teach them, and a smaller part of what they learn from every month. It gets robots started and it rescues them, but videos of people, simulation and practice in the field are what teach them now. The useful part of each remote shift is the failure log (what the robot did, what went wrong, and what the person did to fix it), and the companies that keep it will improve the fastest. Tau already treats its shifts as training. Its privacy page says: "The video our robot records is not a byproduct of the cleaning. It is the training data."

Who sits in the seat

Robot operator is already an old job. Starship's delivery robots have had remote helpers from the start, and in 2019 Kiwibot paid operators in Colombia less than $2 an hour to set waypoints for its robots every few seconds. The new thing is the scale. The number of operators can grow for a while, even as each one runs more robots (1X alone plans 50,000 robots in 2027). And the job changes along the way: first you drive one robot, then you rescue fifty, and then you do the work around the robots (collecting data, running fleets).

Customers will trust the services that say openly how often a person steps in, because robot demos have hidden it before.¹At Tesla's We, Robot event in 2024, Bloomberg reported that people remotely ran many of the Optimus interactions. And Brett Adcock (Figure's CEO) said last year, on the day 1X opened orders for NEO (which a remote operator can take over): "We will not teleoperate them in market". The law is still catching up. In Cologne this month, police are investigating whether someone was steering a robot that touched a 16-year-old. A German law firm explains how it works there. A robot that a person steers is "nur das Werkzeug" (just the tool), so the person is liable. Only an autonomous robot opens a "Haftungslücke" (a liability gap, where nobody is clearly responsible). So for now, a pilot makes the legal question much simpler.

When does the pilot step back?

For basic chores, the pilot steps back by late 2027, earlier than the operators themselves expect. Tau's founder, Alex Koch, plans to phase out his remote technicians "sometime around 2028 or 2029". Føtex's operations director, Simon Kudsk, expects several more years before a robot walks the store fully on its own. And 1X's CEO puts NEO at the level of a human cleaner in 2027, in his best case.

For basic chores, the best case is the right one, because a robot no longer starts from zero. Its brain comes pretrained from the AI labs. Gloves and videos of people show it how to move. And practice in the field fixes the rest (that's how AgiBot reached 95%). Once robots work mostly alone, the ratios move fast too (Pony went from 1:3 to 1:20 in one year). Here are the dates side by side: the companies' own plans, our forecasts and the rules (The labs built the robot brain has the full set of our dates).

After that, the pilot becomes a rescue desk, like Waymo's agents today. Pilots will keep the controls only where their skill or the danger is the point: surgery, bomb disposal, hazardous materials (Minerva builds teleoperated humanoids for that) and shows (Disney's Olaf is driven with a Steam Deck, the handheld games console).

The pilots in Manila watch 50 robots each because the robots barely need them. Every remote-controlled robot is heading the same way, to a pilot who mostly watches, and that's why Minsky's remote-controlled economy arrives just in time to end.