The Best-Selling Vacuum in America Is a Robot

Lex Fridman opened with a number that still sounds like a category error. As of 2019, iRobot had sold more than 25 million robots to consumers: the Roomba vacuum, the Braava mop, and soon the Terra lawn mower. These were not lab demos and not a few dozen units on a warehouse floor. They were machines running unsupervised in other people’s houses. Most robotics companies fail. iRobot, then 29 years old, had not.

Fridman had spent the day at the company’s offices, touring PackBot and the lineage of machines that led to Roomba, Braava, and Terra, then sat down with CEO and co-founder Colin Angle. He released the tour separately and kept this conversation short on origin folklore. What he wanted on tape was how a robotics firm survives long enough to put millions of autonomous machines on carpets, pet hair, and sofa legs.

Midway through, Angle dropped the line that makes the scale concrete. iRobot had, he said, disrupted the entire vacuuming industry. The number-one selling vacuums in the United States were Roombas — not the number-one robot vacuums, the number-one vacuums, counting Dyson, Hoover, and every machine a person still has to push. “That’s really crazy and weird,” he said, and asked for a pause. A robot doing something as essential as vacuuming, as the top-selling object in the category, had started as a crazy idea in a room: do you think we can do this? Let’s give it a try.

Fridman started with Isaac Asimov’s 1942 story “Runaround,” from the *I, Robot* collection, and the three laws: do not harm humans, obey orders, protect yourself. The company is named after that book. So does Roomba follow the laws, and what role should robots play in a world that is still mostly human? Angle called the laws thought-provoking and then took them off the table as an engineering spec. They demand a profound grasp of the world, of the consequences of action, and of a sense of self. That bar, he said, will not be relevant for decades.

If Roomba follows the three laws, and he believes it does, it is not because the robot has intent. It is designed to help people rather than hurt them, to be inherently safe, and to last a long time. “Following the Three Laws is aligned with being a good robot product.” The ethics collapse into ordinary product virtues: safety, reliability, longevity. You do not get there by writing unverifiable principles downward. You get there by shipping a physical object that can run in tens of millions of homes without hurting anyone.

The larger job, he argued, is not optional. Humanity’s average age is rising quickly, and there will not be enough people young enough and spry enough to care for that older cohort. Today’s robots make life a little easier, a little cleaner, maybe a little healthier. In time they will be the difference between a gut-wrenching drop in independent living and a future where people keep control of their days and spend them on what they choose. Angle said he was honored to be on that path. A slightly cleaner floor is the first inch of it.

iRobot had narrowed in on the consumer home: a place that should maintain itself, run healthier and safer, maybe more efficiently. They already vacuumed and mopped; lawn mowing was next. The longer arc, Angle said, is not a menagerie of gadgets but a home that is itself a system, taking an active role in caring for the people inside it. Lights when you leave. A space that quietly does the right thing. Everything they ship is a step on that path.

Fridman summarized the first two steps from the tour. Step one was taking a robot that worked in a lab, or in the field helping soldiers with hard jobs, and making something consumers would own by the tens and hundreds of thousands, that did not break under months of heavy use, and that people liked. Step two was sensing, mapping, and localizing well enough to build a picture of the home that a person could label, giving the robot semantic knowledge of its environment. Those, Angle agreed, were the huge early climbs: from too expensive and too weak to be more than a toy, to affordable and effective enough to create purchase intent.

The step he was excited about ran against his own history. He had spent the life of the company making robots autonomous enough to go do the job without a person. Now he wanted them less autonomous: a partner, not an automaton. If you say you just dropped flour by the fridge in the kitchen, the right thing should happen from that utterance. That is “less autonomous” only in the sense that the robot is listening. It still has to parse context and intent, map the sentence onto its model of the house, and know what to do. You can already tell a Roomba to clean the kitchen and it knows where that is. He called that 1.0.

Arms come after maps, not before. Data had an arm, Rosie had an arm, Robby the Robot had an arm, because fiction assumes a body that does physical work. Angle’s question was sharper: if a robot does not know where anything is, why should it have an arm? Once it knows the kitchen, it may later know the refrigerator; with a hand it might find the handle, open the door, and get itself a beer. A robot bringing a beer while you watch television is, he said, one of the true dreams of the field. The causal chain is easy to miss. The value of a manipulator is set by semantic understanding of space. Without that, a hand is an expensive prop. With it, a hand suddenly has a job. Full autonomy, he added, can only take you so far in a home. The robot should keep a basically right routine and stay ready for a sentence that activates a different set of behaviors.

Fridman listed recent obituaries: Anki, Jibo, Mayfield Robotics with Kuri, Rethink Robotics. Brilliant people, cover-story robots, then gone. Why is it so hard to keep a robotics company alive? Angle answered only partly in jest. Before Roomba he was a high-tech entrepreneur building robots. “It wasn’t until I became a vacuum cleaner salesman that we had any success.” Technology alone is not a business. You have to find a compelling need where the robot delivers clearly more value than it costs. Not a close call you inspect at the margin. Not a plan where you hold your breath and hope. Clearly more value than the price in the store, and in a way that lasts.

Entertainment is even harder. You can be the cat’s meow one year, and 85 percent of toys, regardless of merit, still fail to make a second season. A lot of experiments have asked what the right social companion is, what a home robot should do besides the weekly chores people would rather skip. Angle was not sure the industry had figured out the magical user experience. So you get brilliant roboticists with fascinating machines whose value-benefit equation stays ambiguous, and the business never quite closes.

How big is the space of jobs that actually clear that bar? Angle’s answer was that we are about none of the way to the potential of home robotics, if you keep your eyes open. Roomba is a hugely successful product and still only a foot in the door. He was lucky, early on, that people kept asking when he was going to clean their floor. He had a pile of what he thought were better ideas, “but everyone seems to want their floor clean, and so maybe we should do that.” The good ideas had to earn the right to come next, and they had to match what people actually wanted at a cost the business could stand.

A partnership years earlier with Johnson Wax had taught him a blunt method. They went into homes and watched how people lived, looking for tasks people did not like, did often enough to stay top of mind, and recognized as a burden. Sometimes a society repeats a burden so often that it stops seeing it as removable. You have to sit there and stare. Getting those insights, Angle said, is a lot harder than it seems in retrospect. The implication is not despair. It is that almost none of the housework that could be automated has yet been turned into a business that is better than a person, cheaper than a person, and felt as a burden by the buyer.

People would be unlikely to buy a Roomba that cost $500,000, which is roughly what such a machine would have cost several decades ago. When Angle started, the cost of a robot had a lot to do with the time it took to build. They made them from aluminum. He spent hours in the machine shop on a milling machine, cutting parts one at a time. Then iRobot got into the toy industry, and the worldview cracked. At scale, he realized, you could determine the cost of the robot not by adding up milling hours but by weighing it.

That, he said, is liberating. Three-dimensional CAD, real tooling, and injection molding let you make an arbitrarily complicated plastic part whose cost is basically the weight of the plastic in it. The whole way you think about construction changes. Whether a home robot can exist as a business depends less on the cleverness of the algorithm than on whether you can drop the product into a mature industrial cost function. If you can, cost falls with volume and you get a second and third generation. If you cannot, every unit is a handmade object, you lose money on each sale, and the price never comes down.

Sensing had its own cost function. For a long time Angle tried to invent ways to duplicate human skin, which is an astonishing sensor. He walked off that path. The better question became how many of the problems a working robot faces could be solved with a cheap camera and a large computer. Moore’s law was still running. The phone industry kept handing down better processors and image sensors at consumer prices. About two years before the conversation, he said, machine-vision-capable processors had finally become cheap enough to put on a consumer robot. He had been waiting for that moment.

So iRobot spent years on vision-based navigation and avoided putting lasers on the robots. You could see where the technology trends were going; you should not plant a heavy bet on a transitional part. The formula he wanted was injection-molded plastic, a camera, and a processor that could run machine learning and visual object recognition. Cheap, capable, manufacturable. That, he said, is going to be the future.

Fridman pulled the argument sideways into autonomous vehicles, the other place automation already touches ordinary lives, and into Elon Musk’s line that lidar is a crutch and cameras are the long-term answer. Angle saw huge overlap in the problems and the sensors. Then he inverted the difficulty. His environment, he said, is actually much harder than a car’s. There are no roads. There are T-shirts, steps, and a near-infinite set of patterns, colors, and floor textures. From a visual standpoint a home is infinitely variable. Safety, though, is far easier. His robots are not very heavy and not very fast. If they bump your foot, you think it is funny. Cars have the inverse problem: a more structured world, and a mistake that can kill.

That is why he can say without reservation that vision is the future of the home robot. For vehicles he believes the primary environmental sensor will also be visual, but putting a cheap lidar on as a backup “might not be the worst idea in the world.” The two bets are not the same, so the timelines should not be argued as if they were. He expects at least one iRobot robot in every home. Twenty-five million sold put them in about 10 percent of U.S. homes, a start. Vacuuming, mopping, and soon mowing do not exhaust what a house needs done, if computer vision and low-cost manufacturing keep making new jobs affordable.

A robot that walks around your house and looks at it has a trust problem that can kill the company. Angle said iRobot had leaned way forward because its intended role in the home makes data protection make-or-break. They published a privacy manifesto and promised never to sell user data. They adopted GDPR globally, not only where the law requires it. Images do not leave the robot; visual processing happens on device, and only semantic knowledge of the home goes upstream, with the owner’s consent. They show you what they know. They are trying to use data as an enabler of performance with informed consent, not as a harvest.

The choice is explicit. Let them build a semantic map and know where the rooms are, and you can say “clean the kitchen.” Decline the map and the robot will still do its best job cleaning, it just will not understand that sentence. If you later want it to forget that a room is your kitchen, you should be confident it will. Angle described this as a data-2.0 stance: treat what the robot learns about the house as the customer’s data, with rights attached. Being the good guys, he hoped, would earn the trust required to let robots do more thoughtful work.

Then came the sentence that is more informative than a white paper. Today, when there is a privacy breach, people get very upset, then go to the store and buy the cheapest thing, paying no attention to whether the product honors any privacy standard. “If I put on the package of my Roomba the privacy commitments that we have, I would sell less than I would if I did nothing at all.” That has to change. Earning trust is necessary and not sufficient. What the market needs is a grade-A meat-style standard for privacy that ordinary buyers can understand and actually use in a purchase, so that good behavior is rewarded. Until that exists, every company claims to care, customers cannot tell the difference, and they ignore the dimension. Standards would at least let people start asking who is following them.

Fridman asked whether Angle dreamed of a Roomba with human-level intelligence, the kind you might have a Friday-night chat with after it fetches a beer. On the beer-and-chat half, yes. On human-level intelligence: not in his lifetime. You can still have a meaningful conversation with a robot that has nothing like human intelligence, he said, as long as you remember the conversation is not about making the robot feel good. If the point is to reward the machine, you could put a button on it that makes it smile. If the point is you — how your pet’s day went, how to make the house more efficient, what to know before climbing Everest — that is doable and useful. Robots can help people feel less isolated, learn about the home they live in, and fill lonely gaps with something worth knowing.

His own theory ran the opposite of the usual cartoon. As robots get smarter they will get more emotional, because you cannot survive on pure logic. Only a tiny fraction of situations can be resolved reasonably with logic. Emotion exists because it helps you act under incomplete information, in social life and in competitive or dangerous settings. When robots enter those same situations, something like emotion becomes a requirement, an efficient heuristic evolution already found. People often treat feeling as the opposite of intelligence. Angle treated it as what an agent needs once the world stops being a closed-form problem.

Last question: if he could spend a day with a science-fiction robot and pick its brain, whom would he choose? Data, from *Star Trek*. Data is smart, has been through a lot trying to save the galaxy, and would have plenty to say about emotion and machines. If Angle could ask Data one question and get a deep, honest answer, it would not be about silicon. It would be, “What’s Captain Picard really like?” A man who spent 29 years putting vacuuming robots into homes wanted to use a fictional android to learn more about a person. That is a fair summary of his stance. The robot does not need to become a human. It needs to leave more room for the human parts of a life.