The Kelly Distinguished Lecture on Robots and Jobs features preeminent scholars in fields of significance to robotics. The visiting lecturers, in addition to presenting seminars on topics relevant to robots in the workplace, participate in informal discussions with Georgia Tech faculty and students.
The inaugural lecturer, Dr. Rodney Brooks, presented “The Case for More Robots” on Friday, March 11, 2016. Brooks is the founder, chairman, and chief technology officer of Rethink Robotics, Inc.
Henrik Christensen: It gives me great pleasure to introduce Rodney Brookes for the first Kelly Distinguished Lecture on Robots and Jobs. For those of you who don’t know the history of this lecture series: The sponsor is Clint Kelly, who’s up here in the front. Clint has a very distinguished career. He was on the strategic computing program at DARPA, and he was part of the first autonomous land vehicle program. So he has a very strong history of interest in robotics and advanced computing and these kinds of technologies. Clint then moved on to be the senior CEO for SAIC (http://www.saic.com/), and now Clint travels the world in search of penguins and soon he will have a book coming out on that and that will be very nice. [Clint Kelly travels the globe photographing different species of penguins for a future book publication.] So Clint and I talked about this, and he said he’d love to do a lecture series about robots and societal impact. So he’s very kindly sponsored the series, which is co-organized by me and Nancey Green Leigh, from Architecture. [Now, the College of Design.] And I’m very much interested in finding out what the overall societal impact is from this. And I’m very pleased we have Rodney as the first speaker in this lecture series. As Magnus [Magnus Egerstedt] said, “We got the most famous roboticist in the world to show up for the first lecture, which is absolutely amazing.” And Rodney’s famous in so many ways. Rodney has his roots in Australia, and then he came to Stanford. And of course he came up with a model-based vision system. Every vision system has to have a short name, so Rodney called it ACRONYM. Why call it anything else when you can call it ACRONYM and be done with the letter puzzle?
He moved on to a number of academic positions and ended up at MIT. He had a very distinguished career there, and ended up running CSAIL, to be very successful. Rodney also spawned iRobot, and now he’s doing Rethink Robotics, so I think Rodney’s done it all in terms of fantastic research, really transitional stuff. I’m very pleased that Rodney is here and that he would do the first lecture for us on robots and jobs. So, please welcome Rodney Brooks.
Rodney Brooks: I’m not sure I would’ve done the lecture if I’d known I was doing the first one [laughs], but I guess someone has to do it first. Also I want to say that I’m not sure I know much about this. Most of this talk is new, I’ve never given it before, and we’ll see how successful or not it is. Oh by the way, that’s Mildred. Mildred works at a plastics factory in Connecticut. I don’t know her education history but she’s worked in that factory on the line for about 25 years, so I’m assuming she doesn’t have above a high school education. This was an hour after she’d met her first robot ever, and she’d already programmed the robot to do a task, there in the background. I think the question people are worried about with robots and jobs is whether we’re going to get more of this: the rich on one side and the poor on the other, and nothing for the poor. I think that’s a complex question, I think there’s a complex set of things happening in the world right now that make it hard to make predictions. I’m more of the flavor that those of the left (poor) will need more robots to raise their standard of living, so we’ll look at that.
People talk about the holding out of the middle class, and there’s certainly an aspect of that. This is data from 2001 to 2011. These are the tenths of the wages. This is the lowest tenth, next tenth, the next, and blue; it’s the average wage in that percentile in 2001, and in red, it’s the average wage in 2011. And you can see on the left things have gone down a bit, but when you get up here, these guys on the right are doing a whole lot better. On the left, the scale is job growth, and you can see they’ve been growing negatively in the middle, growing on the left. And there’s a little bit of growth in these really rich people here.
When I looked at this data, it’s not as bad as I thought from the press reports. It’s not good, but it’s not as bad as I thought. There’s certainly a trend there. Now, how do robots play into this? Henrik mentioned that I founded iRobot, which makes the Roomba. And 16 million of those are out there. When I came to the U.S. in 1977, there were three mobile robots in the world. So, it feels good to have built 16 million of them, I wish 16 billion, but oh well. But, these robots get built in China. This is a photo from 2005, and these are people building robots. And you see them packed together pretty well. They’re doing fairly simple tasks with simple tools.
I spent a lot of time in factories in China, as we were building up the Roomba. But also in my role as director of CSAIL, I worked with a lot of Taiwan-based electronics manufacturers that operate in China. Companies like Quanta Computer, which makes lots of the world’s servers, lots of things for Dell, Google, etc. They were investing in research at MIT because they wanted to stop being just manufacturers and start developing new products and become brands, which many of them had. So I spent a lot of time in Taipei with the founders of those companies who’d been running them for thirty years. And around 2004, they started telling me stories like the following. “Back in the 90’s, in mainland China, if we wanted more workers, we’d put up a sign this big sign saying, ‘labor needed.’ The next morning, there’d be a line all the way around the block. And now we have to advertise for workers.” That’s 2004. At our factories in China, golden week will turn out to be a disaster because everyone would go home, and a lot of them wouldn’t come back. So we’d go down from two shifts, to one shift. Back in 2004, we’d started to see that maybe the infinite supply of labor in China wasn’t going to be there forever. This is a low-end factory, and this is a higher-end factory. You can see how packed people are, close to each other. They’re working side by side. And in both of these sorts of factories, if it’s a well-run factory, the turnover is 16% per month, which is very high. So labor is very hard to get in China.
Since 2004, it’s just gotten worse and worse. We’re not going to have this infinite supply of labor in China. Here’s one of the reasons by the way. This is from the CIA World Fact Book. This is the population distribution in 2014 of China in five-year age groups. Notice what’s happening here. The number of 19-year-olds is dropping dramatically. This is certainly an effect of the one-child policy. If you look at the number more closely, 19-year-olds are going down by 30% in absolute numbers in a ten-year period. So, there’s just less people to do the work.
I just went on the web and grabbed a story. This is saying China reversed the one-child policy around December of last year because they see this drop of population, and loss of economic vitality. So they’re trying to increase the number of workers by going back to two children allowed. In the U.S. we see an aging population also. The number is in thousands in all industries in manufacturing. Fortunately, it’s about one tenth so you can just compare. There are a million twenty to twenty-four-year-olds in manufacturing. But for younger ages, it’s below one tenth. When you get up to here, it’s a lot more than one tenth in this 45 to 54 year range, and a little over in the 55 to 64 year range. Across all industries and jobs, there are a lot of older people out there doing a lot of things pretending not to be older people. There’s an aging of the U.S. manufacturing workers’ population compared to general workers.
Here’s some data. This is the shortage in highly trained manufacturer talent. Six hundred thousand U.S. manufacturing jobs remain unfilled, and 82% report a moderate or serious shortage in skilled production workers. And the average age of highly skilled workers in manufacturing in the U.S. is 56. Why do people not really want to go into these jobs? We don’t see why people are going to jobs in agriculture and working in the fields. We don’t see such a line for people waiting to work in manufacturing either in the U.S. or in China.
I’m about to have a political joke in the next slide, and it’s only 67% accurate. So avert your eyes if you’ll be offended. This is something I found on Facebook… It’s actually only two of the three wives. But I think this captures the point.
People don’t want these bad jobs. They’re not attracted to them. I was giving a talk around commencement activities at Brown three years ago. I was talking about robots and one of the undergraduates said, “You’re taking jobs from people with these robots.” And I said, “Would you like your son to work in a factory?” and he said, “Oh, not my son, not my son.”
So people don’t aspire to work in factories. You don’t get parents that say “I hope my child grows up and gets a good job in the factory.” That’s not what people aspire for. They want a higher-level job, a more cognitive job. That’s what people aspire to. When you don’t have as much immigration to fill the factories there’s not as many people. This country and other countries are voting fairly heavily against more immigration. So there’s a shortage of people. And there’s a shortage of people in China for manufacturing. So maybe in the long term robots in manufacturing are going to hurt jobs, but it’s hard for me to get overexcited about taking away jobs from people right now. But, I saw this happening in 2004, I thought well what can we do, I’m the robot guy, and we’ll put more robots in factories. And this is what modern factories with robots look like, and you’ll notice one thing there. There ain’t no people. There are little cage places that the people can go into near the robots, but the robots are far too dangerous. In a body shop it’s just all robots, no people. They’re doing very similar tasks again and again. And in a modern production line, the actual chassis may change and they may adapt the motions preprogrammed, but it’s very heavily choreographed. Worse, the capital cost of buying the robot is only about a fifth the cost of installing the robot. You’ve got the robot; you’ve got this ancient controller box, and they build brand new things with controller boxes still.
You know, I bought an ABB robot for my lab at MIT around 2007, and I realized it was possible to control the controller by Ethernet, but there was no way of plugging Ethernet in so I called up ABB, and I said, can I get the Ethernet thing, and they said “Oh, that’ll be about 3000 dollars, to have an Ethernet cable.” So they’re in the dark ages, and they’re making a lot of money. And then you need to program the robot, if you want cameras, you buy separate cameras, you install them, you have a separate computing system, and then you connect the computing system on there with PLC (programmable logic controllers), which come from 1968, roughly. You have to get electricians to put those cables in; it’s a complex sort of thing. And then you have to put cages around them, either light cages, or physical cages, so a different person goes in through this black cage here, it’ll stop the robots, because they’re just not safe.
That means that the placement of helping workers with robots is you have to build a whole system where people and robots can’t mix. It didn’t seem like the Chinese factories or U.S. factories could have these robots in there without throwing everything away and having incredibly high capital cost, for bringing a whole robot system. Which is what automobile companies do in the body shop and paint shop, they’re all robots no people. But if you go to final assembly it’s almost no robots. It’s all people, and no robots. And they’re incredibly hard to use, and that’s very different from what we’ve seen in our consumer devices, from a Roomba, desktop computers, laptops. These things are easy to use. We don’t need multiple trays to install them.
By the way, who amongst you has a smart phone? And who amongst you went to community college to learn how to use that smartphone? Okay. To use those robots, you’ve got to go to community college and get courses to program. You’ve got to have a degree.
Consumer technology is easy to use. If you made the robots easy to use, that would be one thing. If we made them safe for humans to be around that would be another thing, and then maybe we could just have robots and people together. And a whole bunch of things happen. So it’s not just robots its industrial Internet, it’s additive manufacturing, cheap sensing, cheap computation, and basic connectivity. These are all things that the robot industry hasn’t used. So we started to think about how to put those together. And our first robot is Baxter, and it’s working side by side with a person. One of the big problems, as we show pictures of our robots working with humans, is everyone is scared people will see what they’re doing, and the robot will copy them. So these are blurred here. These are putting pieces in a heat-sealed box for a phone. And you can look at the color of the blurred thing there and figure out which phone it is. Here are robots working side by side with people. Creating the really cheap plastic furniture that you buy at K-Mart or Target that you keep in the basement and don’t let anybody see. And we have Baxter, our first robot, and Sawyer is our second robot. Baxter’s been out about 3 ½ years and Sawyer’s just been out a few months, and just is really only getting into mass production. These robots operate in these factories. One thing that’s difficult about them is there is no universal gripper that works, so every installation people make their own grippers. We provide kits. A lot of people use 3D printers, even the little tiny places…the little tiny factories, 20 people. They use suction, they use electric grippers, and they have their robots, in amongst people. This guy has his back to the robot. Don’t do that with an industrial robot. And it’s not because the robot has intent; it just doesn’t care. Here’s another heat sealing. These women are putting things into these fixtures, they’re going into the heat sealer, and the robot takes them out when they’re hot, and drops them off.
This is our new robot, Sawyer. It’s bigger than a person, but it’s meant to go in a place where a person goes. It’s got longer reach with its arm, because it doesn’t have hips—a person has hips. This is Shiwei. One of our engineers showing it in zero-course gravity compensator mode, moving it around, to show how easy it is the get it over somewhere. It’s got a Cognex camera in the end of its arm, and it all comes integrated. You train the robot by showing it things in the world, but they’re not in fixed coordinates, they’re in local coordinates—multiple local coordinate systems. And we tried to design the interface so that a worker who doesn’t have a degree in computer science, has never written a line of code, can get the robot to do interesting things, by showing it pieces of the task and using a graphical user interface and using buttons that are on the robot. People say, “Why don’t you have a touch screen up there?” Well, most factory workers wear gloves all the time, so touch screens don’t work. They say, “Why do you have those stupid eyes there?” or “Why does it have that silly grin on its face?” The eyes are there because when it’s about to reach somewhere, it glances ahead of reaching, so the people around it know what it’s about to do. You don’t have to train them that that’s what it’s about to do, they just pick up on that. If a person was going to reach somewhere, they’re going to glance before they reach, so it’s hardwired to take that cue. We try to find cues and portray them without getting into the uncanny valley, too human like.
This is Sawyer. No lines of code were written here, this is a machine-tending task. She’ll turn that piece of metal there and then pick up the top and put it back, using force sensing to locate them, so then she’ll have to do two reaches to fill the machine. The person who operates the machine has to close the door, so the robot closes the door, and then the person presses the go button, so that’s what the robot does. So we tried to make it so that the person can show it a task in the way that a person does a task, and have the task operate, but without any lines of code. No coding at all. So in the interface, we tried to make it so that we have two slogans for it: “It knows what you mean and it does what you want” and “Simple things are simple to train. Complex things are possible.”
I often look at systems where you have to know so much to do the simplest thing. I saw a programming interface for an Android phone a couple of years ago. You had to go to 14 different screens to get “hello world” to work. Simple things were not simple. So we tried to make it really simple. We have this graphical user interface on the screen that shows a map of where the two arms are, what it can reach. Here’s how you show the robot how to do something. Here’s a robot arm, here’s a human arm. There’s the parallel jaw gripper. The button on the side changes the state of the gripper. Press it once, it opens; press it again, it closes. So I bring the arm over here, bring it down there, I press the button, the fingers close. Immediately, the robot looks up, a little icon, and says, “Oh, they must want me to pick something up here. Because I had nothing in my hand, I closed my fingers, now there’s something in my hand. This must be a pick up.” And we could have a menu saying “What do you want to do here, it’s going to be a pick up, oh how do I pick up? I close my fingers,” but instead, we just infer that. Suppose this was a cup, suppose the invisible arm is being trained by my arm and I bring it over with the fingers closed, put it in the cup, press the button, and they open, now it feels some force there so it says, “Oh, this is a pick up, and how I pick up here is I open my fingers.”
So it’s just lots of little tiny pieces of inference which then build up a task description you can go in and click on it, and say “Oh, I want to train the object and put vision on it, I want to show it what to do.” This is showing what an object is. It gives you a little set of shadings on the object, saying, “Which one looks the best, am I seeing the object, am I segregating it from the background?” so that’s how it chooses which algorithm to use for vision. It’s not like giving a demo in a research lab where we prepare before the visitor comes and make sure the demo works. It’s got to work out of the box, and has to work in the factory everyday, and the people who are programming the robot know nothing about computer vision. One of the big things we got when we first started going out to the factories was, “Why do I have to do a training of the robot for what the object looks like. Can’t it see what the object looks like?” The academics or engineers know, “of course we have to train the robot.” So we have to make it as easy and intuitive [as possible] to use. And there’s a lot of technology around—I’m going to come back to this later.
So we use a series of elastic actuators. Some of you here may have seen, or were at, the DARPA robot challenge run by Gill Pratt…Gill Pratt and Matt Livingston, an old friend of Charlie’s [Charlie Kemp] graduate students. So here, series elastic actuators lets you do force sensing at each joint. So here, he’s moving it around, and here’s the magnitude of the force of each joint. This joint doesn’t have any gravity load on it, so it’s pretty small, but this joint has got a lot of gravity force on it but it’s sensing force using Jacobian, and you can get what the resulting force is in any coordinate system. The robot is able to sense forces everywhere and that’s important. That lets us interact with the world where we don’t know very precise precisions. It’s going to pick up that thing and put it in a well. And every time it picks it up, the person is going to put it in a different location. So it’s going to pick it up, not knowing exactly how far along that boundary. So it goes down, slides until it feels a force, puts it down. It gets very high accuracy because the well is the same size as the object. Here it is in real life. This is a steel case factory. It’s putting metal in a press brake. It’s sliding it in against the stops that are in there for a person.
There’s a pile of pieces of metal near the left arm of the robot. It doesn’t know where they are. The person just rolls it up and they’re a few centimeters opposite. So there are a few centimeters of uncertainty. The robot reaches down, uses suction to grab it, but then aligns it with the backstops in the press brake. So it’s precise enough. This is a steel case; this will be in drawers of office furniture. It picks it up, not really knowing where it is, but it slides it in using PLC. We were thinking of having a foot on it at one point to press the GO button, to go on to the next. Here, it’s putting a circuit board into a circuit board tester; I’ll come back to these later. But it’s using force to see, and again, no programming here. Just the person has shown it the strategy of how to seat this, and it goes in there and reliably seats it to about a 200-micron positioning. And you’ll noticed what’s happening if there’s something in the way. It realizes it’s not working. No one wrote any error handling. The error conditions are put in the software stack. It tries again, seats it. Here’s something you really don’t want to try at home with your industrial robot. Put your hand in the way of this thing, comes down, feels that it’s not quite right, backs up without crushing the person’s fingers, which is important. So it’s safe to be in close with this robot, and be able to use the same sort of fixtures that people use.
Here is, in practice, having local coordinate systems, which in the user interface you can attach these local coordinate systems—one there, one over on the conveyor—bring the robot over. It finds the initial landmark, and then has the local coordinate systems around the global coordinate systems. So to us, or you guys, in academia, that sounds pretty straight forward, but to have that in an actual product is a big innovation. I often say, “We’re trying to bring robots into the late 20th century in manufacturing.”
VIDEO: At our facility at Steelcase, we do architectural walls, and office furniture, primarily steel products. We process about 14 tons of steel a day, about a million square feet under our roof. So we go from blank sheets of steel to formed, welded, painted product in about three days. So we’re being driven to find more ways to implement automation and advanced technologies to take us to the next level. Our demographics here at Steelcase, we have an older, mature workforce, so what we’ve done is really put robotics and new technology right in their hands, and let them be able to see it, touch it, feel it, and then we have a better understanding of how they’re trying to utilize a Baxter or a Sawyer within the workforce. The biggest difference between a Sawyer and an industrial robot would be the programming. It was impressive how easy it was to program the machine for pick-up points and placement points, and it hit the mark the way it was supposed to. We will be implementing Sawyer in areas where it has a highly repetitive task or mundane task, to help free up our operators to do more important functions in their work cells. So Sawyer is going to be picking up a plate and loading it into a fixture. And then he has to go back and grab the tube, and load it into the fixture, and there are two fixtures that eventually he’ll load. It’s a high volume part for us, and what we want to do is to be able to utilize this piece of equipment on an off shift, and we can just utilize more operators to do different things depending on the schedule that we’re trying to run. We don’t want employees to feel that the robot is there to take their job away. It’s there to assist them with their job. We’re looking at employee engagement, employee enhancement, and being an employee multiplier.
RODNEY BROOKS: Okay, so what that was showing there was putting a square plate—putting a leg—these are legs of tables, four of them, and then that machine rotates, and the welding happens on the other side. That’s how the person used to operate that welding machine. You put the piece in the thing and the welding is automated on the other side. So now they’re using Sawyer to do that at night when there’s no one else around in the factory, and just produce lots and lots of these things, stack them up, so that they can be used in an assembly that is more complex, and more dexterous, and the person has to do it. So they’re enabling a second shift. When I talk to small factories, and Steelcase is not a small place, but people in the U.S. do not want to have a second shift. It’s very hard to get workers for second shifts. It’s very hard to get trustworthy workers and people don’t really want to work overnight in a factory. It’s a really hard labor problem.
So our customers talk to us about how we get rid of a worker, how we move some task that is not very high value—putting those pieces in the machine, a very dumb sort of task—and have the robot do that so we can use more of the people we do have for more complex tasks. Is that disingenuous? I don’t think so, but that’s what they say to us. So now, that’s what we’ve been doing with robots in factories, but here’s what’s really happening in the world. We see it with older workers in factories. There’s a demographic conversion happening. For you, young people here, you’ll have a lot of old people slowly coming along. Just a lot of old people, I’m sorry. “So, I’m inventing good robots,” that’s what I said. This is Japan. This is in 2015, and this is what it’s going to look like in 2045. Over here, this is the 50 to 64 year age group; it’s going to be down by half. Look at the 65+, it’s already over a quarter, it’s going to be a third of Japan that is going to be over 65. That’s a lot of people. But it’s not just Japan. Japan is the leading indicator, but it’s true in other parts of the world. It’s true in China; it’s true in Europe. This is the percentage of people, working age or older, who are working. It’s gone down a little bit over 40 years. It’s about to go down, a whole bunch more as we get an aging population. Even in China, which still has most people working, it’s going to go down comparable amounts by 2045. It’s going to be a lot of old people, and not so many younger people. A big portion of the population will be retired or in elder care and who’s going to look after them as they get more and more frail? There will be a lot of them. Well, I think technology can help.
Does anyone know what this is? It’s a Mercedes S Series, and I tend to think of this as an elder care robot. And why is it an elder care robot? Because it has a lot of driver assistance features. It lets the elderly drive longer, and safer. And in the U.S., certainly, if you don’t have a car you’ve lost your independence. So it gives people independence longer, by driver assistance. I don’t think we are going to see general or completely autonomous driving cars in amongst our roads as quickly as some other people think. We’re going to see more and more over the next ten years—more and more driver assist vehicles. That is an example of the older people that can afford them, and the younger people that can’t, but they will trickle down. That’s an example of older people getting an advantage from robotics. Letting them drive independently longer. But, eventually, they will need physical therapy. It’s the old being cared for by the moderately old. And that’s also a trend to attack. There’s not enough to look after the truly elderly, but we’ll see a transition from how elderly are helped today with passive devices, which let people bathe [themselves] longer and have their dignity and control of their lives, to more active devices, like Charlie Kemp’s group here in biomedical engineering [is working on], building some of these sorts of devices which will help the old. I don’t think we’ll see these out in products in the next five or so years, but a lot of research [will be] happening over the next ten or fifteen years for this to get out and help [people]. Fortunately, it will be at just the perfect age for me to need those machines, so thank you Charlie. And I don’t think they are going to look like that.
There are other things in robotics that are helping the elderly. As the elderly lose mobility, they tend to shop online more. Certainly, the next generation who has been using computers all their lives will shop online more and more, which leads to, in general, how do you get all the things in the package? And a lot of it is manual, still today, but there are robotics systems [in place]. This is the Kiva System. It was bought by Amazon. It moves these shelves around, moves them up to the person who does the picking, so it keeps the person from running around.
We’ve seen the robot that people get from Symbotic now for case stacking. So fulfillment is another thing with robotics. To provide a service to the elderly who are not able to get out and shop anymore. And you’ve got to get the stuff to the house. How do you get the stuff to the house? People have started talking about drones. Give that last little piece of delivery. Is it going to be drones? I’m not sure, but here’s a company, funded by the Skype founders, which is a little device that is supposed to go the last three kilometers, I think, with the articles for the house and bring them from distribution centers all around town and get them out to the house so you don’t have to have people drive around all of this stuff because we’re not going to have enough labor. There’s just not enough labor. I think we also want robots in our houses, as people get older to let them have more independence. To let them move independently, but I see three major problems. The mobility problem, the messiness problem, and the manipulation problem, and I wish there was more research in these areas because I think this is going to be important for healthcare in a fifteen-year time frame. There will be great numbers of companies who are going to make a lot of money by building systems and services.
This is what real houses look like. That little four-wheeled thing coming from the founders of Skype, gets to the house, uh oh, it can’t get up the stairs, unless you cut to another scene and R2D2 is up the stairs. But in real life, it’s tricky getting up and down the stairs in the house. How do you do that? Well, there are a few platforms out there that can do it. There’s a PackBot from iRobot, but you don’t want that to go up and down those stairs, trust me. There are a few more, but don’t let them fall over, because that’s the end. And this is the older version of the Atlas. We just saw new versions in videos over the last couple of weeks. But, I don’t want a thing full of hydraulics in my house, I’m sorry. So, I think we need research on mobility to get around houses where it steps up, steps down, steps into the houses, good solutions for that.
This comes from some research project’s simulation of what a kitchen is like where they’ve got simulated robots operating this kitchen. Aaron Edsinger, one of my Ph.D. students, had this in his Ph.D. defense. He showed his kitchen, and it’s a little different. This is what real life is like. Real life is more like that. So, how do we deal with that messiness? And here, there’s good news. The good news is the gaming industry has made 3D sensors. So now when you go to a vision conference or robotics conference, you see a large number of papers in low-cost 3D sensors dealing with that messiness and clutter. I saw in many labs today, people are using them here. This is common in academia. So I think the messiness is sort of on the way to getting solved. People are working on it. A lot more needs to be done, but it’s happening. Then there’s manipulation. Leonardo da Vinci knew hands were really special [da Vinci created the “Study of Hands”]. Our robot hands haven’t changed in 40 years. There’s just parallel jaw, grippers, and suckers. The Kiva system I showed you before brings the shelves with goods over to the person, then the person uses their hands to do it. Amazon, who bought Kiva, had the Kiva picking challenge in May of last year at the robotics conference to try and stimulate people: “How to pick stuff up,” it’s not a solved problem at all. A lot of the problems come from that hands are complex devices. They can’t just do one thing; it’s a team effort. So we made Baxter available, the research robot, and I was hoping, and some people are, that someone would work on new hands for Baxter, because they don’t have to worry about the arm part. I think there’s a bunch of things that you have to do to make a hand. You have to make progress simultaneously on mechanisms, on sensors, on materials, and on the algorithms—all four at once. So this is a real team effort, and this is something that I’m hoping that groups in academia will decide is worth doing, this is worth making better hands for robots, because there will be an impact across everything in robotics.
So that’s robots, and where we are with robots. But there are also megatrends, which I think have an even bigger influence on the world than robotics does, so I just want to bring up a couple. One is urbanization. In the next 40 years, we have to build 200 mega cities in the world to handle all the people who want to move to cities. That’s as much building as we have right now and building—I think that the carbon footprint of doing the construction is 30% to 40% of the carbon in the world. So tremendously large scale operations—where does the labor come from? Aged people? And how do we deal with the carbon footprint? This is going to change our economy worldwide in a major way; two hundred new cities… By the way, global warming… Global warming is going to cause the dislocation of 200 other cities. We already saw that in Hurricane Sandy in New York. We’re seeing cities getting flattened now. Another city, Venice, just had its highest floods ever last week. In other places, in deltas and valleys, Bangladesh, we’ll see more and more flooding and more destroyed cities. That’s a lot of infrastructure, a lot of construction.
I don’t know how these are going to play with robots and employment. I think that it’s a complex problem. I want to finish up, and just talk a little bit about what we see in the press, which argues about the rational version of robots in employment. But then we see people like Stephen Hawking, Elon Musk, saying we’ve got to worry about these robots, they’re going to be super intelligent, and they’re going to take over from us, and there’s no role left for people. I just have one word. “Relax; okay? Just relax,” but I want to explain why I think people are making this error. None of them are in AI, none of them are in robotics, but they keep repeating this and the press loves it. There has been great progress in AI, recently driven by Pete Norvig. Fantastic progress in learning speech, images, but it’s overhyped by our science. And why is it overhyped?
In the person, the person has a high performance at some task. We have a natural understanding of how that performance generalizes. So we see that performance and we go oh, if a person can do that task well, they must be able to do a whole lot of tasks around it. But that’s not true for these algorithms. The same generalization doesn’t work. I think that’s the fundamental mistake that people are making. We understand in a human, what human competence is, but we don’t understand machine competence. So let me show you some examples. Many of you in computer science have seen these. This is the standard image labeling, and you know, man in black shirt is playing guitar is a pretty good label for this image. These are automatically produced and they’re not perfect…Young girl in pink shirt is swinging on a swing. It looks like it’s a swing set and she’s grabbing it, not swinging on it. If your robot was a rescue robot it may make a mistake that way. Boy is going backward on a wakeboard? No, boy is actually on a little trampoline on the beach in the water. It’s not perfect. But they all work by—this is the Google example—going through a deep neural network to get out labels for images. They’re actually pretty good. But remember, why does Google want this? Google wants this so you can search using words to find images. It doesn’t have to be perfect, and it’s not perfect; it sort of works in a way that is not how we work. This is the probability—0.9% chance of being a person, 60% chance of being a person. That’s 0.9% chance it could be a person? We know it’s not a person given the rest of the content. You put it all in, and you can reverse it. This is again from Stanford, and it’s looking for soccer ball, soccer ball, soccer ball, tennis ball. We know that’s a tennis ball. And this is one of the Google examples. These are the five top things. It’s a laptop. It’s a hair drier—why could this be a hair dryer? Well, [look at] the eyepiece of a camera near her hair. Some other things…a seatbelt. It is a saw, the seatbelt, but it’s not quite human-like vision. I don’t think any of you here would say she’s using a hair drier. The policewoman out next to her police car, and she’s drying her hair.
Here’s something from NYU. It has deployed more vision processing than anyone in the world because it now processes 2 billion images a day on Facebook. But here, they’ve got this image labeling, that successfully labels this image—all these images. So they said how much noise do we have to add to break it? The image labeler fails on every image on the right. No person fails on them. It’s not intuitive what’s going on. It’s a little bit of noise, and it breaks the network's fact propagation and it doesn’t label these, whereas it labels these. Though, if we see a generalization, if we try to generalize and say, okay, it labeled these, therefore it can see the world like we can. No, it doesn’t see the world like we can. It’s looking at an image. Here’s another group. This is a trained deep neural network that says, yeah, that’s a guitar with 98.9% confidence, and I’m 99.9% confident that’s a penguin. And then they used genetic algorithms based on what Karl Sims, as some of you might remember from the early 90s, generates synthetic images that would trigger these sorts of recognitions in this trained deep mind. And it says yeah, that’s a guitar, I’m 99.99% sure that’s a guitar. So the genetic algorithm went and looked for something to make it trigger this network. That’s a penguin, that’s a baseball, that’s a matchstick. You can see we wouldn’t make those mistakes. You can probably see why it thinks this is a school bus, why it thinks that is a computer keyboard, but we wouldn’t make those mistakes. When I was talking to Peter Lee, the vice president of Microsoft, running research there, they previously worked with DARPA. He said, “Oh, when I saw these results I called up DARPA and said we’ve got to worry about people putting up fake signs that the driverless cars will see and obey even though the people walking around won’t notice that their signs are going to trigger the vision algorithms in the cars. Well, maybe that’s going to be we’re getting attacked in a new way. These things around town look innocuous, but they will make you just drive into a river. So getting back to this… When we have a human with a particular performance, we can deduce their competence. We cannot deduce the generalized competence of a robot or a machine. The next image is a joke of mine; this is me labeling an image. Some of you will get it. Just from yesterday, here we have the human “Go champion” being beaten by a deep-learning network and some people say that’s the end of humans. Now, Go can be played by robots. On my Facebook feed someone said it would be interesting to know whether if we gave that program and that human a 29 by 29 Go board, who would win. It wasn’t obvious to all my friends on Facebook that we don’t know how it’s going to generalize. And so the fact that people say this is a great advancement for AI…I don’t think it’s such a great thing. So I think we can relax about a lot of stuff that’s happening in the world. I think, as for the robots, we’re going to interact with them closely, they’ll be safe to be around, they’re easy to use, and they will look after me in my old age, I hope. Thank you.