Interview with Rodney Brooks
ROBOTS (Sabine): Hi Rod, welcome to Robots.
Rodney Brooks: Hi. Thanks for having me.
ROBOTS: It is great to be here in the offices of Rethink Robotics. Congratulations on the launch of your new product.
Rodney Brooks: Thank you.
ROBOTS: Can you tell us about the vision behind Rethink Robotics?
So this idea of outsourcing low cost labor did not seem sustainable
Rodney Brooks: Yeah, I have done a lot of manufacturing for iRobot in the Far East, and so I have seen a lot of Chinese factories, and I realized there were a whole bunch of problems with manufacturing in the Far East. If you are manufacturing consumer products, you have to make decisions 30 weeks ahead of when they are going to hit the shelves, and all the retailers in the US want pack outs, so you have to decide how many of each product for each retailer and sometimes those retailers have gone out of business before the stuff arrives. Also I was noticing that it was getting harder and harder to get employees in China because the standard of living is going up, and this has happened many times in the history of outsourcing manufacturing from Europe and from North America since the Second World War. Originally, we outsourced to the low cost region of Japan until their standard of living went up. Then we moved to South Korea, then we moved to Taiwan, then we moved to Southern China and even in the late 90s some things, when I was building toys and stuff in China, were too expensive to do there already and had gone to Vietnam. So this idea of outsourcing low cost labor did not seem sustainable, so I started thinking how could we make manufacturing in North America, or in Europe, or even in Japan more attractive, and being a roboticist of course, the answer was robots. That is the only answer that I can come up with for any problem. But the idea was to make robots that could work in small factories for low cost goods because industrial robots up until now have largely been in auto factories. In the US at least 70% of all industrial robots are in automobile factories and the hundreds of thousands, 300,000 small manufacturers in the US, hardly have any robots at all because they are so hard to introduce into the factory and so expensive to introduce. So I wanted to build a low cost robot that a small factory could have, and that had to be programmable by an ordinary factory worker. So it is a little like 30 years ago, we went from the mainframe to the PC, and with the mainframe computer, ordinary office workers had no access to computation, but with the PC they did. So our robot, Baxter, is meant to be a robot for the ordinary factory worker to interact with.
ROBOTS: Tell us a bit more about Baxter. What does it look like, what are its sensors and actuators?
Rodney Brooks: Well Baxter, and this did not start out this way, but Baxter ended up looking a bit like a human, with human form. It has 2 arms and a head – I will talk about the head in a minute – the 2 arms are slightly longer than a normal person’s arms, because we wanted to approximate human reach, but Baxter does not have any hips. As people, we bend with our hips in order to reach things. Baxter cannot so Baxter has longer arms, looks a bit like an Olympic swimmer in that sense. It has got 2 arms; it is on a pedestal. The pedestal is not mobile by itself. You wheel the robot around and just lock down the wheels wherever you want Baxter to work. The 2 arms have force sensing in every joint. They have 7 degrees of freedom. They have various buttons and knobs on the arms so you can interact with the robot and tell it things. And then where there would be a head, there is an LCD screen. The LCD screen always tries to point towards you if you are near the robot, so you can see what is on it, and by default it has a pair of eyes. Now people say why do you have these graphic eyes on the robot? When it is working, just as a human does, it glances where it is about to reach to, so if you are a person near by the robot, you are never surprised by its action, because it always looks where it is about to reach so you sort of get this cue ahead of time, which helps to make it safer and predictable. But then when you are teaching the robot a new task, training the robot, the eyes move off to the corner and you get to see inside Baxter’s brain, and you interact through some menus, with the knobs on its wrist and on its forearms, and select various tasks for it to do. It has cameras, one in the LCD screen that looks out at people, two in its chest that look at the work space, and one in each wrist. So if you want to teach it about a certain object that it is going to interact with, you grab the wrist, it goes into a sort of floating mode, the arm, you bring the wrist above the object, and now you say ‘look at this, this is the object.’ Up on the screen appears what it is seeing through its hand camera. It [displays] its best guess of where the object is separating it from the background, tries to figure ground separation with no external input from the user, and you can say ‘yeah, you sort of got that or no, you have not got it.’ And if you tell it it’s got it, then the robot starts moving the arm around from different angles, using its range sensor that is in its hand, getting multiple views of the object to build a model of it, that can then [be used to] recognize it, for whatever task you are about to have the robot do.
ROBOTS: Based on this description, we sort of get a sense of the flexibility of the things that we can do with this robot. Can you help us understand who you think might be maybe the top 3 professions or the top 3 people who would be using this to start out with?
Rodney Brooks: Well it is really about what sort of workplaces it is. So we have not tried concentrating on any one particular industry. We have more had the idea that it is going to be useful for small manufacturers who do not normally have robots. To have a normal industrial robot right now it typically takes somewhere like 8, 10 or 12 months to get the robot installed from the time you decide you want to have a robot, to purchasing it, to going through the systems integration, to relaying out your factory floor so you have safety cages around the robot, etcetera. It take many months, often more than a year and the cost of the installation can be 3 to 5 times the capital cost of the robot. Our goal was 1 hour. 1 hour from when you decide you want a robot, or the robot shows up, to having it do useful work, and we have done that many times in tests. We have driven up to a factory that we have never been into it before and within one hour we have had the robot up on the factory floor doing a useful task that was already being done in the factory. And in a few minutes, we train the line workers how to train the robot. They have never seen a robot before. They do not have to understand 6 dimensional vectors or quaternions or anything like that. They just grab the robot arm, show it objects, select things from the menus and train the task. So within that framework, small manufacturing plants, we have a series of tasks that we are getting Baxter to do. Baxter is a physical platform but there are going to be software releases every few months, and every few months there will be new sorts of tasks. So what it can do right now is essentially material handling, picking things up off conveyor belts or putting things on conveyor belts, putting things in arrays, in single layer boxes, moving stuff around. That it is pretty good at. Then by the time we get to January it will also be able to press buttons on machines to do testing of objects. It knows about conveyor belts right now. It will know about some other objects early in 2013, in particular cardboard boxes. It knows a little bit about cardboard boxes right now, but during next year it will start to understand cardboard boxes to a much better degree, multiple layers where you pack things in one layer in a box you put a separator on you pack in a next layer. It cannot do that right now with its current software, but it will be able to do it later next year. And then machine tending is an important thing. Putting things into fixtures, taking them out of fixtures, doing tests, that sort of thing.
ROBOTS: What other uses are there for Baxter other than manufacturing?
Rodney Brooks: We have concentrated on manufacturing and our whole software system, the way it is shipping to our first customers, is that. But we also realize that this price for the robot, with two arms, it is so much cheaper than any other platform out there that we think it is going to be great for researchers, and it is based on ROS, the robot operating system. So we are scrambling right now to get together an SDK for the robot, and we expect in the January time frame to start shipping it to researchers, all over this country at least, and, with an SDK, for people when they start doing all sorts of interesting things. Some people will work at a sort of AI-ish level of tasks. Other people might work down at the control level with it. But I think it is going to be interesting to see when we have lots of people out there do different things with the platform and I think they will start going into areas that are not manufacturing at all. One of the things I would like to see is going to eldercare. It is impossible I think; it does not make sense today to go to venture capitalists and say fund the robotics company for eldercare, but I really believe we are going to need more robotics in elder care in the next 20 to 30 years because of the aging baby boomers, and just not enough people to help them in their old age. So by having a low cost platform out there it lowers the barrier to the entry to start doing experiments in that arena for instance. So I expect to see lots of researchers do wild things with this robot that we would never have thought of.
ROBOTS: As the general public and roboticists start reprogramming this robot, are you worried about questions of liability?
Rodney Brooks: The robot is fairly inherently safe, but if you reprogram almost anything you can make it do bad things you know, unsafe things, so there has got to be some common sense there and we are going to have to rely on researchers having common sense.
ROBOTS: What would you say are the main limitations right now?
Traditional robots have had super human precision, our robot is much more like a person
Rodney Brooks: Baxter is not a precise robot, and we did not design it to be a precise robot. We designed it to be adaptable, so that if something moves in its environment it uses its cameras to figure out where it is and go grab it, flexible in that it will know about many things and therefore you can train it easily to do particular tasks, because it knows most of the tasks already it is just the specifics it needs, and ease of use so that any person, anyone can train the robot to do a task, and a very low cost robot also. So they are the metrics that we evaluate whether we are successful or not. And it has to be safe to interact with people, to be safe we have to make it compliant, by making it compliant it is not stiff, and then accuracy is not what you get. But a human arm is not accurate either. When we do a task, we do not precisely place a little tiny screw in an iPod by having our arm reached out, extended 60 centimeters away from our body, and try and just do it by a position. We do it by sensing. We look at where the tip of the screw is, and then we actually usually brace our hand with our smaller fingers so that our thumb and index finger can then precisely move the screw relative to the visual sighting of where the hole is. So humans do not have precision. Traditional robots have had super human precision, our robot is much more like a person but it senses force in every axis, and that is much more like a human. So it tries to do tasks in human like ways.
ROBOTS: One of the limitations is that it currently has two fingers. How do you see extensions happening in the manipulation area?
Rodney Brooks: Manipulation and dexterity is I think the great frontier for lots of robotics. Baxter actually, by default, has nothing at the end of its wrists. In the end of its wrist it has a camera and a range finder and then a mechanical plate and a connector. We are currently shipping two different sorts of hands for it. One is a two-fingered parallel jaw-gripper with two fingers that just slide together back and forth and the other is a suction gripper. You can attach either of those. The two fingered gripper by the way comes with a whole hand kit which has multiple sets of fingers and multiple pads that you can put on the inside or outside of the fingers and also guards to stop how the hand moves. So there are about a hundred configurations of those fingers but we expect to bring out new versions of the hand, over time, which can be installed, and we expect third parties to design their own hands. We expect larger customers to design hands in house, for some particular things, but we would love it if some little companies started building dexterous hands for Baxter, and sold them, and we will be publishing the interface if that can be done.
ROBOTS: You spoke a lot about the importance of human robot interactions. What did you take form your previous research at MIT to apply to Baxter?
Rodney Brooks: To me the most important thing was the idea of authenticity, that whatever the robot promises by its appearance, either just its physical appearance or the way it acts, that it is authentic in what it is promising; it delivers on those promises. I have said a few times, if you build a robot that looks like Albert Einstein, it ought to be as smart as Albert Einstein, or people are going to be disappointed. So we try to make Baxter be authentic in all the cues it gives. It has eyes, graphical eyes, but we do not try to make them do things that it is not delivering on. If it is glancing somewhere that means it is going to reach there. That is a pretty simple thing to pick up on, and so people can know that if it glancing somewhere it is really paying attention to that thing and it is going to do something with that space. When you are training it, if it does not understand you, its eyes give a confused look. Well that is because it does not understand you, but it is not saying it is sad, it does not have tears, it does not do things that would be inauthentic. It is surprised, its face turns red if you get really close to it, just to sort of give you the clue that maybe you are too close to the robot. But it is authentic in what it tells you. So authenticity I think is the big lesson I learned from our research. If it is not authentic, people are going to get very frustrated with it very quickly
ROBOTS: Do you have an example of an interaction with Baxter gone bad and did you learn anything from that?
Rodney Brooks: Oh and here is an interesting thing, as engineers we make certain assumptions about what is obvious. That may not be what the end users see as obvious. We would spend a lot of time in factories talking to a lot of end users; we thought we understood things and we came up with different characteristics of different sorts of users – the tinkerer, the person who just wants to follow a script, etcetera. But the big surprise the first time we took it out to a factory, where the task involved an object that had to be trained, the thing that the line workers just thought was weird was they would show it where to pick stuff up from, they would show it where to put stuff. But then they would say – and our interface at that time had a much different feel to what it has now – ‘Why do I have to show it the object? Why do I have to train the object? Can’t it see the object? I just showed it the object. I moved the object around.’ But we had a very separate place where – ‘Now this is the object.’ – you know putting the object in an object library, and that just did not make sense to an end user. To an engineer or a computer vision person or a robot scientist, ‘Oh, of course you have to have the object library; that is obvious.’ But it was not obvious to the end user, so we have tried to deemphasize that and make it more automatic, so that happens without the person really having to think about that.
ROBOTS: How difficult is the vision problem?
Rodney Brooks: As a start up, we have to only choose to solve problems that are solvable. The general vision problem is not solvable, so what we do is recognize objects by a single view so if there is an object that looks different upside down, that is 2 totally different objects, as far as Baxter is concerned. It is just appearance based, and it is essentially 2D, so we do not get a 3-dimensional reconstruction of the object. It is views from particular angles. Now fortunately most things that you want to deal with in a factory would tend to have 1 or 2 or 3 stable positions that it is going to be on a conveyor belt so you just … if you show it how to grasp each of those then it has learnt the objects. Now Baxter does have specialized vision routines for special sorts of objects, so for instance it can see a conveyor belt in the distance just by detecting motion, and it knows conveyor belts are this sort of continuous area of motion in the same direction, and so it knows about conveyor belts. In the first version of the software, there is nothing special about boxes but in development we got software that just sort of ‘Gong! That’s a box; that’s a box; that’s a box.’ It sees boxes, because cardboard boxes are something that you can pull out and build a special detector for it. And also, again not in the first version, but face detection software in the camera above the screen. So there are certain sorts of things which people can work on, and face detection is one of those where we have got better at over the last few years by intense work, but generic object recognition, or understanding flexible material, is hard. Flexible material, transparent material is another hard one for us so there are some limitations
ROBOTS: With so many sensors and modeling and planning, do you feel like you are still able to stay true to your initial philosophy of behavior based robots?
Rodney Brooks: Oh, very much. It is behavior-based robot. In fact when you train it to do a task, there is no representation of the sequence, of the task, and that is good. Instead there are representations of what to do. When you train it you are essentially teaching it some behavioral rules – ‘When you see this thing over here pick it up.’ – but then it knows, those behaviors, like the behaviors that control the hand, know some common sense things. You cannot put anything down if there is nothing in your hand. You cannot pick something up if there is already something in your hand. So, with those simple behaviors and simple rules, if the robot picks something up and it is moving to put it somewhere in the box and someone comes up and grabs the thing out of its hand, it does not go with its empty hand to the box and mimic putting it there, it realizes ‘There is nothing in my hand; I cannot put anything in the box.’ It will try and go and get something else. If it is putting something down, you have said put it down on the table here, but, as it is moving it down, it feels a force, it figures well there is something else there and I am just going to let go of the object now. If you have told the robot to fill a box with 4 by 3 widgets, 4 across and 3 deep and there are 6 in there and it is going to put the 7th one in but a person comes up and fills in 3 extra ones in transit, it realizes that. It does not try to fill the full holes. So it has got all its behaviors and it responds to the world and it responds to changes in the world through those behaviors.
ROBOTS: Baxter was 4 years in the making. Is that short or is that long? What were some of the challenges along the way?
Rodney Brooks: Yeah, the Baxter that we see today is really 6th generation that we built during that time. The robot is low cost and is made in North America. So we were really looking for ways to build a low cost robot all the while. We did not start off ‘Let’s build the best possible robot, then figure out how make it low cost.’ We started off with low cost as a design goal, so we did a lot of experimentation in different materials, and we use some interesting materials in order to bring the cost down and the production down. So for instance, there are 2 major sorts of gearboxes in the arms – the large gearboxes and the small gearboxes. The large gear boxes are made of die-cast metal, but the small gear boxes are plastic, with glass impregnated; it is a very low cost very light weight material, very easy to mould, and we worked on it quite a while to get to be accurate enough that we do not need to do any post-machining for those gear boxes, so that makes it a very low cost gear box structure. The gears themselves, instead of being cut from metal – and we experimented with plastic gears at one point; we could not quite make the work well enough – they are powdered metal. So it is powdered metal, compressed and then heat-treated, much, much cheaper than conventional gears, but not as good as conventional gears, so we have very interesting computational models of the gears, that are running at full speed to counteract the imperfections in the gears, so that we can get by with low cost physical stuff and computation dealing with the inadequacies of the low cost materials to make the performance better.
ROBOTS: So the challenge is really integration; getting the software, the hardware, getting the cost down, putting all these things together to finally get one product
Rodney Brooks: Very much integration. Even manufacturing the robot, there are sub assemblies built in 12 different states in the US and they all come together. So the last few months, because we are building this in mass production, the last few months a lot of our engineers have been working on test equipment that is now spread all across the US, in these various small factories that are building parts of the robot, that do quality tests and calibration of sub modules and those calibration numbers get stored inside sub modules and that gets pushed through the whole assembly line, so we could not have just built one robot at low cost. It is because we are going after mass production, but we had to do a lot of work to make mass production work and it is complex. With 2 hands on, it has 18 gearboxes in the robot, 7 different gear sets and 18 gearboxes. That is a lot of mechanical stuff. It has a quad core Intel main processor but it has got over a dozen embedded processes connected by Ethernet within the robot. So there is a lot of computation, a lot of complexity there to make it all work.
ROBOTS: Revolutionizing manufacturing is a very ambitious goal. When will you know that you have reached a turning point?
Rodney Brooks: Well we are at the juncture right now. ‘We have decorated the restaurant; we have trained the wait staff; we have cooked the first meals; now are the customers going to like the food?’ We will see, I think in 6 months we will know whether we have got something that is plausible, because we will have enough experience with enough customers to know if plausible or not plausible, but then it will be another couple of years before we see whether it really catches on whether it is beyond plausible and ultra desirable.
ROBOTS: What could be possible failure points? Is it a question of mentality of the people or the hardware or…?
The relevant question for our robot is how easy is it to train
Rodney Brooks: One of the challenges for us is that anyone who has thought about industrial robots before their first question is ‘What is the accuracy and repeatability of the robot?’ And that is the wrong question. So, to give you the counter for that, we had a little small factory in Connecticut, owned by an individual, this factory, actually his father had built the factory and now the son was running the factory, and he knew about industrial robots, he had never had one before. We brought the robot there. We showed it how to do a task, got his workers to show it how to do a task, it was doing the task and he came over, and then one of our engineers came and just pushed the robot 10 centimeters to the side. And he was, ‘Oh no, now we have got to go through it again.’ But the robot just kept working, because it was using vision. So we moved the base 10 centimeters, it did not affect the robot operating. So that shows you how precision and repeatability is the wrong question, but nevertheless that is the question everyone asks, and they want to compare it to other industrial robots and it is not a good comparison. I try to explain, for instance, that Honda with their new ASIMO robot, their newest one, they are very proud of it because it can actually jump on one foot and it can jump a few millimeters. And I say to people, would you ask one of the big industrial robot manufacturers how high their robot can jump? No, you would not. It is the wrong question. It is the right question for the Honda ASIMO. It is not the right question for traditional industrial robots and the questions for traditional industrial robots are not the right questions for us. But that is a real dangerous place for us, because people will continue to ask that question and we have to figure out how to let them see that it is no longer a relevant question, not a relevant question for our robot. The relevant question for our robot is how easy is it to train, how adaptable is it to changes in the environment, which by the way a traditional industrial robot scores terribly on but those are not questions that they are normally asked about them.
ROBOTS: You are clearly one of the best-known roboticists and you have been very successful in business between iRobot and Rethink Robotics. What are your tips for young entrepreneurs who want to start their own business?
Rodney Brooks: Well I was very successful with iRobot, and now Rethink is on the verge of either … we are about to find out. But I have also had 3 failed companies, and what I did not understand earlier on was you have to build something that people want to buy. And I think often a researcher is so in love with their particular research technique, and they are so used to selling it to funding agencies. ‘Here is this great technique; fund me to do more of it.’ They think ‘Oh, if I start a company with this technique, it is the technique I am going to sell.’ You cannot sell a technique. You have to sell something that provides value to the customer, so they want to buy it, and that is, I think, the hardest lesson for researchers to get.
ROBOTS: Would you advise them to go for a low-hanging fruit or a transformative goal like you are doing now?
Rodney Brooks: It depends on your personality. We see the same thing in web companies. Some of them have gone for low-hanging fruit and done very well, and then there have been transformative ones, like Facebook or Google. You are not going to produce an enormous company unless you try something transformative, but the risk goes way, way up so it depends on your appetite for risk and your willingness to deal with failure, because most transformative ideas fail.
ROBOTS: How is the funding landscape right now for robotics?
Rodney Brooks: Robotics is doing well right now, especially in the Boston area. We have had a big success with iRobot and then the Kiva system, which has its Swiss connections, was sold to Amazon for a large price, so that has got people realizing that robot companies can do real things. And there are a lot of robot companies in the Boston area, so the Boston venture capitalists are sort of taking it as a point of pride, ‘We are going to keep pushing robotics here.’
ROBOTS: Last question. What is next for Rethink Robotics?
Rodney Brooks: Well Rethink Robotics has to produce in scale. We have to be producing thousands and thousands of robots and selling thousands and thousands of robots in order to be the sort of success we want. We need to listen carefully to our customers, adapt the robot if we see that it needs changes, and then think about other parts of manufacturing that Baxter cannot do, and see if we can come out with some follow on products that will work in those areas.
ROBOTS: Do you already have any projections on sales 10 days after the launch or…?
Rodney Brooks: We have no projection on sales and the example that I use for that is the Roomba. At iRobot, when we first announced it, in September 18th 2002, exactly 10 years before Rethink’s Robot, the board thought there was no way we could sell 15,000 robots before Christmas. Well, we sold 70,000. So making predictions, we would have been totally off base with the predictions we had. I cannot tell you how many Baxter robots we are going to sell at this point. We really need I think a good year or so in the field to get a feel for how well it is accepted and how useful people find it before we can come up with any sort of accurate models.
ROBOTS: Excellent. Thank you very much, Rod, for being here with us on Robots.
Rodney Brooks: Thanks for having me again.