科林·安格爾：iRobot

The following is a conversation with Colin Angle. He's the CEO and co founder of iRobot, a robotics company that for 29 years has been creating robots that operate successfully in the real world. Not as a demo or on a scale of dozens, but on a scale of thousands and millions. As of this year, iRobot has sold more than 25 million robots to consumers, including the Roomba vacuum cleaning robot, the Bravo floor mopping robot, and soon the Terra lawn mowing robot. 29 million robots successfully operating autonomously in real people's homes, to me is an incredible accomplishment of science, engineering, logistics, and all kinds of general entrepreneurial innovation. Most robotics companies fail. iRobot has survived and succeeded for 29 years. I spent all day at iRobot, including a long tour and conversation with Colin about the history of iRobot, and then sat down for this podcast conversation that would have been much longer if I didn't spend all day learning about and playing with the various robots and the company's history. I'll release the video of the tour separately. Colin, iRobot, its founding team, its current team, and its mission has been and continues to be an inspiration to me and thousands of engineers who are working hard to create AI systems that help real people. This is the Artificial Intelligence Podcast. If you enjoy it, subscribe on YouTube, give it five stars on iTunes, support it on Patreon, or simply connect with me on Twitter at Lex Friedman, spelled F R I D M A N. And now, here's my conversation with Colin Angle. In his 1942 short story, Runaround, from his iRobot collection, Asimov proposed the three laws of robotics in order, don't harm humans, obey orders, protect yourself. So two questions. First, does the Roomba follow these three laws? And also, more seriously, what role do you hope to see robots take in modern society and in the future world? So the three laws are very thought provoking and require such a profound understanding of the world a robot lives in, the ramifications of its action and its own sense of self that it's not a relevant bar, at least it won't be a relevant bar for decades to come. And so if Roomba follows the three laws, and I believe it does, it is designed to help humans, not hurt them, it's designed to be inherently safe, and we designed it to last a long time. It's not through any AI or intent on the robot's part. It's because following the three laws is aligned with being a good robot product. So I guess it does, but not by explicit design. So then the bigger picture, what role do you hope to see robotics, robots take in what's currently mostly a world of humans? We need robots to help us continue to improve our standard of living. We need robots because the average age of humanity is increasing very quickly, and simply the number of people young enough and spry enough to care for the elder growing demographic is inadequate. And so what is the role of robots? Today, the role is to make our lives a little easier, a little cleaner, maybe a little healthier. But in time, robots are going to be the difference between real gut wrenching declines in our ability to live independently and maintain our standard of living, and a future that is the bright one where we have more control over our lives, can spend more of our time focused on activities we choose. And I'm so honored and excited to be playing a role in that journey. So you've given me a tour. It showed me some of the long histories, now 29 years that iRobot has been at it, creating some incredible robots. You showed me Pacbot. You showed me a bunch of other stuff that led up to Roomba, that led to Braava and Terra. So let's skip that incredible history in the interest of time, cause we already talked about it. I'll show this incredible footage. You mentioned elderly and robotics in society. I think the home is a fascinating place for robots to be. So where do you see robots in the home? Currently, I would say, once again, probably most homes in the world don't have a robot. So how do you see that changing? What do you think is the big initial value add that robots can do? So iRobot has sort of, over the years, narrowed in on the home, the consumer's home, as the place where we want to innovate and deliver tools that will help a home be a more automatically maintained place, a healthier place, a safer place, and perhaps even a more efficient place to be. And today, we vacuum, we mop, soon we'll be mowing your lawn. But where things are going is, when do we get to the point where the home, not just the robots that live in your home, but the home itself becomes part of a system that maintains itself and plays an active role in caring for and helping the people live in that home. And I see everything that we're doing as steps along the path toward that future. So what are the steps? So if we can summarize some of the history of Roomba, you've mentioned, and maybe you can elaborate on it, but you mentioned that the early days were really taking a robot from something that works either in the lab or something that works in the field that helps soldiers do the difficult work they do to actually be in the hands of consumers and tens of thousands, hundreds of thousands of robots that don't break down over how much people love them over months of very extensive use. So that was the big first step. And then the second big step was the ability to sense the environment, to build a map, to localize, to be able to build a picture of the home that the human can then attach labels to in terms of giving some semantic knowledge to the robot about its environment. Okay, so that's like a huge, two big, huge steps. Maybe you can comment on them, but also what is the next step of making a robot part of the home? Sure, so the goal is to make a home that takes care of itself, takes care of the people in the home, and gives the user an experience of just living their life and the home is somehow doing the right thing, turning on and off lights when you leave, cleaning up the environment. And we went from robots that were great in the lab, but were both too expensive and not sufficiently capable to ever do an acceptable job of anything other than being a toy or a curio in your home to something that was both affordable and sufficiently effective to drive, be above threshold and drive purchase intent. Now we've disrupted the entire vacuuming industry. The number one selling vacuums, for example, in the US are Roombas, so not robot vacuums, but vacuums, and that's really crazy and weird. We need to pause that. I mean, that's incredible. That's incredible that a robot is the number one selling thing that does something. Yep. Something as essential as vacuuming. Yep. Congratulations. Thank you. It's still kind of fun to say, but just because this was a crazy idea that just started, you know, in a room here, we're like, do you think we can do this? So, hey, let's give it a try. But now the robots are starting to understand their environment. And if you think about the next step, there's two dimensions. I've been working so hard since the beginning of iRobot to make robots are autonomous, that, you know, they're smart enough and understand their task enough, they can just go do it without human involvement. Now what I'm really excited and working on is how do I make them less autonomous? Meaning that the robot is supposed to be your partner, not this automaton that just goes and does what a robot does. And so that if you tell it, hey, I just dropped some flour by the fridge in the kitchen, can you deal with it? Wouldn't it be awesome if the right thing just happened based on that utterance? And to some extent, that's less autonomous because it's actually listening to you, understanding the context and intent of the sentence, mapping it against its understanding of the home it lives in and knowing what to do. And so that's an area of research. It's an area where we're starting to roll out features. You can now tell your robot to clean up the kitchen and it knows what the kitchen is and can do that. And that's sort of 1.0 of where we're going. The other cool thing is that we're starting to know where stuff is. And why is that important? Well, robots are supposed to have arms, right? Data had an arm, Rosie had an arm, Robbie the robot had an arm. I mean, robots are, you know, they are physical things that move around in an environment and they're supposed to like do work. And if you think about it, if a robot doesn't know where anything is, why should it have an arm? But with this new dawn of home understanding that we're starting to go enjoy, I know where the kitchen is. I might in the future know where the refrigerator is. I might, if I had an arm, be able to find the handle, open it and even get myself a beer. Obviously, that's one of the true dreams of robotics is to have robots bringing us a beer while we watch television. But, you know, I think that that new category of tasks where physical manipulation, robot arms, is just a potpourri of new opportunity and excitement. And you see humans as a crucial part of that. So you kind of mentioned that. And I personally find that a really compelling idea. I think full autonomy can only take us so far, especially in the home. So you see humans as helping the robot understand or give deeper meaning to the spatial information. Right. It's a partnership. The robot is supposed to operate according to descriptors that you would use to describe your own home. The robot is supposed to, in lieu of better direction, kind of go about its routine, which ought to be basically right, and lead to a home maintained in a way that it's learned you like, but also be perpetually ready to take direction that would activate a different set of behaviors or actions to meet a current need to the extent it could actually perform that task. So I got to ask you, I think this is a fundamental and a fascinating question, because iRobot has been a successful company and a rare successful robotics company. So Anki, Jibo, Mayfield Robotics with their robot curry, SciFi Works, Rethink Robotics, these are robotics companies that were founded and run by brilliant people. But all, very unfortunately, at least for us roboticists, all went out of business recently. So why do you think they didn't last longer? Why do you think it is so hard to keep a robotics company alive? You know, I say this only partially in jest that back in the day before Roomba, you know, I was a high tech entrepreneur building robots. But it wasn't until I became a vacuum cleaner salesman that we had any success. So, I mean, the point is technology alone doesn't equal a successful business. We need to go and find the compelling need where the robot that we're creating can deliver clearly more value to the end user than it costs. And this is not a marginal thing where you're looking at the scale and you're like, yeah, it's close. Maybe we can hold our breath and make it work. It's clearly more value than the cost of the robot to bring, you know, in the store. And I think that the challenge has been finding those businesses where that's true in a sustainable fashion. You know, when you get into entertainment style things, you could be the cat's meow one year, but 85% of toys, regardless of their merit, fail to make it to their second season. It's just super hard to do so. And so that's just a tough business. And there has been a lot of experimentation around what is the right type of social companion, what is the right robot in the home that is doing something other than tasks people do every week that they'd rather not do. And I'm not sure we've got it all figured out right. And so that you get brilliant roboticists with super interesting robots that ultimately don't quite have that magical user experience and thus that value benefit equation remains ambiguous. So you as somebody who dreams of robots changing the world, what's your estimate? How big is the space of applications that fit the criteria that you just described where you can really demonstrate an obvious significant value over the alternative non robotic solution? Well, I think that we're just about none of the way to achieving the potential of robotics at home. But we have to do it in a really eyes wide open, honest fashion. And so another way to put that is the potential is infinite because we did take a few steps, but you're saying those steps are just very initial steps. So the Roomba is a hugely successful product, but you're saying that's just the very, very beginning. That's just the very, very beginning. It's the foot in the door. And I think I was lucky that in the early days of robotics, people would ask me, when are you going to clean my floor? It was something that I grew up saying, I got all these really good ideas, but everyone seems to want their floor clean. And so maybe we should do that. Yeah, your good ideas. Earn the right to do the next thing after that. So the good ideas have to match with the desire of the people and then the actual cost has to like the business, the financial aspect has to all match together. Yeah, during our partnership back a number of years ago with Johnson Wax, they would explain to me that they would go into homes and just watch how people lived and try to figure out what were they doing that they really didn't really like to do, but they had to do it frequently enough that it was top of mind and understood as a burden. Hey, let's make a product or come up with a solution to make that pain point less challenging. And sometimes we do certain burdens so often as a society that we actually don't even realize, like it's actually hard to see that that burden is something that could be removed. So it does require just going into the home and staring at, wait, how do I actually live life? What are the pain points? Yeah, and getting those insights is a lot harder than it would seem it should be in retrospect. So how hard on that point? I mean, one of the big challenges of robotics is driving the cost down to something that consumers, people would afford. So people would be less likely to buy a Roomba if it cost $500,000, which is probably sort of what a Roomba would cost several decades ago. So how do you drive, which I mentioned is very difficult, how do you drive the cost of a Roomba or a robot down such that people would want to buy it? When I started building robots, the cost of the robot had a lot to do with the amount of time it took to build it. And so that we build our robots out of aluminum, I would go spend my time in the machine shop on the milling machine, cutting out the parts and so forth. And then when we got into the toy industry, I realized that if we were building at scale, I could determine the cost of the Roomba instead of adding up all the hours to mill out the parts, but by weighing it. And that's liberating. You can say, wow, the world has just changed as I think about construction in a different way. The 3D CAD tools that are available to us today, the operating at scale where I can do tooling and injection mold, an arbitrarily complicated part, and the cost is going to be basically the weight of the plastic in that part, is incredibly exciting and liberating and opens up all sorts of opportunities. And for the sensing part of it, where we are today is instead of trying to build skin, which is really hard. For a long time, I spent creating strategies and ideas around how could we duplicate the skin on the human body because it's such an amazing sensor. Instead of going down that path, why don't we focus on vision? And how many of the problems that face a robot trying to do real work could be solved with a cheap camera and a big ass computer? Moore's law continues to work. The cell phone industry, the mobile industry is giving us better and better tools that can run on these embedded computers. And I think we passed an important moment maybe two years ago where you could put machine vision capable processors on robots at consumer price points. And I was waiting for it to happen. We avoided putting lasers on our robots to do navigation and instead spent years researching how to do vision based navigation because you could just see where these technology trends were going. And between injection molded plastic and a camera with a computer capable of running machine learning and visual object recognition, I could build an incredibly affordable, incredibly capable robot. And that's going to be the future. So on that point with a small tangent, but I think an important one, another industry in which I would say the only other industry in which there is automation actually touching people's lives today is autonomous vehicles. What the vision you just described of using computer vision and using cheap camera sensors, there's a debate on that of LIDAR versus computer vision. And the Elon Musk famously said that LIDAR is a crutch that really in the long term, camera only is the right solution, which echoes some of the ideas you're expressing. Of course, the domain in terms of its safety criticality is different. But what do you think about that approach in the autonomous vehicle space? And in general, do you see a connection between the incredible real world challenges you have to solve in the home with Roomba? And I saw a demonstration of some of them, corner cases literally, and autonomous vehicles. So there's absolutely a tremendous overlap between both the problems a robot vacuum and an autonomous vehicle are trying to solve and the tools and the types of sensors that are being applied in the pursuit of the solutions. In my world, my environment is actually much harder than the environment an automobile travels. We don't have roads. We have t shirts. We have steps. We have a near infinite number of patterns and colors and surface textures on the floor. Especially from a visual perspective. So the way the world looks is an infinitely variable. On the other hand, safety is way easier on the inside. My robots, they're not very heavy. They're not very fast. If they bump into your foot, you think it's funny. And autonomous vehicles kind of have the inverse problem. And so that for me saying vision is the future, I can say that without reservation. For autonomous vehicles, I think I believe what Elon's saying about the future is ultimately going to be vision. Maybe if we put a cheap lighter on there as a backup sensor, it might not be the worst idea in the world. So the stakes are much higher. The stakes are much higher. You have to be much more careful thinking through how far away that future is. Right. But I think that the primary environmental understanding sensor is going to be a visual system. Visual system. So on that point, well, let me ask, do you hope there's an iRobot robot in every home in the world one day? I expect there to be at least one iRobot robot in every home. We've sold 25 million robots. So we're in about 10% of US homes, which is a great start. But I think that when we think about the numbers of things that robots can do, today I can vacuum your floor, mop your floor, cut your lawn, or soon we'll be able to cut your lawn. But there are more things that we could do in the home. And I hope that we continue using the techniques I described around exploiting computer vision and low cost manufacturing that we'll be able to create these solutions at affordable price points. So let me ask on that point of a robot in every home, that's my dream as well. I'd love to see that. I think the possibilities there are indeed infinite positive possibilities. But in our current culture, no thanks to science fiction and so on, there's a serious kind of hesitation, anxiety, concern about robots, and also a concern about privacy. And it's a fascinating question to me why that concern is amongst a certain group of people is as intense as it is. So you have to think about it because it's a serious concern. But I wonder how you address it best. So from a perspective of vision sensors, so robots that move about the home and sense the world, how do you alleviate people's privacy concerns? How do you make sure that they can trust iRobot and the robots that they share their home with? I think that's a great question. And we've really leaned way forward on this because given our vision as to the role the company intends to play in the home, really for us, make or break is can our approach be trusted to protecting the data and the privacy of the people who have our robots? And so we've gone out publicly with a privacy manifesto stating we'll never sell your data. We've adopted GDPR not just where GDPR is required, but globally. We have ensured that images don't leave the robot. So processing data from the visual sensors happens locally on the robot. And only semantic knowledge of the home with the consumer's consent is sent up. We show you what we know and are trying to go use data as an enabler for the performance of the robots with the informed consent and understanding of the people who own those robots. We take it very seriously. And ultimately, we think that by showing a customer that if you let us build a semantic map of your home and know where the rooms are, well, then you can say clean the kitchen. If you don't want the robot to do that, don't make the map. It'll do its best job cleaning your home. But it won't be able to do that. And if you ever want us to forget that we know that it's your kitchen, you can have confidence that we will do that for you. So we're trying to go and be a data 2.0 perspective company where we treat the data that the robots have of the consumer's home as if it were the consumer's data and that they have rights to it. So we think by being the good guys on this front, we can build the trust and thus be entrusted to enable robots to do more things that are thoughtful. You think people's worries will diminish over time? As a society, broadly speaking, do you think you can win over trust not just for the company, but just the comfort that people have with AI in their home enriching their lives in some way? I think we're in an interesting place today where it's less about winning them over and more about finding a way to talk about privacy in a way that more people can understand. I would tell you that today, when there's a privacy breach, people get very upset and then go to the store and buy the cheapest thing, paying no attention to whether or not the products that they're buying honor privacy standards or not. In fact, 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. And that needs to change. So it's not a question about earning trust. I think that's necessary but not sufficient. We need to figure out how to have a comfortable set of what is the grade A meat standard applied to privacy that customers can trust and understand and then use in their buying decisions. That will reward companies for good behavior and that will ultimately be how this moves forward. And maybe be part of the conversation between regular people about what it means, what privacy means. If you have some standards, you can say, you can start talking about who's following them, who does not have more. Because most people are actually quite clueless about all aspects of artificial intelligence, the data collection, and so on. It would be nice to change that for people to understand the good that AI can do. And it's not some system that's trying to steal all the most sensitive data. Do you think, do you dream of a Roomba with human level intelligence one day? So you've mentioned a very successful localization and mapping of the environment, being able to do some basic communication to say, go clean the kitchen. Do you see in your maybe more bored moments, once you get the beer, to sit back with that beer and have a chat on a Friday night with a Roomba about how your day went? So to your latter question, absolutely. To your former question as to whether a Roomba can have human level intelligence, not in my lifetime. You can have you. I think you can have a great conversation, a meaningful conversation with a Roomba without it having anything that resembles human level intelligence. And I think that as long as you realize that conversation is not about the robot and making the robot feel good. That conversation is about you learning interesting things that make you feel like the conversation that you had with the robot is a pretty awesome way of learning something. And it could be about what kind of day your pet had. It could be about how can I make my home more energy efficient. It could be about if I'm thinking about climbing Mount Everest, what should I know? And that's a very doable thing. But if I think that that conversation I'm going to have with the robot is I'm going to be rewarded by making the robot happy, well, I could just put a button on the robot that you could push and the robot would smile. And that sort of thing. So I think you need to think about the question in the right way. And robots can be awesomely effective at helping people feel less isolated, learn more about the home that they live in, and fill some of those lonely gaps that we wish we were engaged learning cool stuff about our world. Last question. If you could hang out for a day with a robot from science fiction, movies, books, and safely pick its brain for that day, who would you pick? Data. Data. From Star Trek. I think that A, data is really smart. Data has been through a lot trying to go and save the galaxy. And I'm really interested actually in emotion and robotics. And I think you'd have a lot to say about that. Because I believe actually that emotion plays an incredibly useful role in doing reasonable things in situations where we have imperfect understanding of what's going on. In social situations when there's imperfect information. In social situations, also in competitive or dangerous situations that we have emotion for a reason. And so that ultimately, my theory is that as robots get smarter and smarter, they're actually going to get more emotional. Because you can't actually survive on pure logic. Because only a very tiny fraction of the situations we find ourselves in can be resolved reasonably with logic. And so I think Data would have a lot to say about that. And so I could find out whether he agrees. If you could ask Data one question, you would get a deep, honest answer to what would you ask. What's Captain Picard really like? OK, I think that's the perfect way to end it. Colin, thank you so much for talking today. I really appreciate it. My pleasure.