This article and video are part of the TechXchanges: ROS: Robot Operating System and Women in Science and Engineering (WISE), as well as the Then and Now library series. You can also check out more TechXchange Talks videos.
Building robots isn't an easy task. One of the platforms that makes the job easier is the Robot Operating System (ROS). Open Robotics supports ROS as well as Gazebo, a robot simulation environment.
I spoke with Open Robotics' Katherine Scott, Developer Advocate at Open Robotics, about ROS and her work in the robotics arena, as well as what she has seen develop over the years.
Links
Transcript
Wong: Robots are a lot of fun and they've been going on for quite a while.
Katherine, you've actually been working with robots for probably almost your whole career. Could you tell us when you started working with robots?
Scott: Well, so I started working with robots in probably about 2001 when I started undergrad. And it's been kind of an off and on thing all the way through.
I kind of tend to focus on computer vision, but always that usually leads back to working with robots. I started with Open Robotics coming up on three years ago, but I've known them basically from the start. I was friendly with a lot of people at Willow Garage prior, you know, the original creators of ROS and Gazebo or ROS and OpenCV.
A couple of years ago, they kind of approached me and said, “Hey, would you be interested in working here?” I'm like, “heck, yeah.”
So I live in San Francisco, and I've never been really interested in working in the South Bay and Silicon Valley. I said the only place I'd ever work at in Silicon Valley is Open Robotics. They finally came and said they want me to work in Open Robotics and we'll let you stay in San Francisco. That works for me. I'm sold. Excellent!
Wong: Well, could you tell us a little bit how the field of robotics has progressed since you started with it? What kind of things have you seen.
Scott: Well, it's changed a lot. Right. It's changed from early 2000. It was big, if you think about it. There's a big wave in the eighties and nineties. A very, very specific, very, very regimented robot that did like one thing. They did one thing. They did it fast. They did it well. They cost a lot of money. They took a while to install. You had a lot of subject matter expertise to do it.
That was the situation for a very long time and what we've seen, particularly in the past, I don't know, let's say five to six years, is that things have changed a lot. Robots have gotten cheaper. They've gotten easier to use and they've gotten way more adaptable to like real-world situations.
So, for example, you've gone from the sort of example in the eighties and nineties that you had a car or (you were) building a car, you got a robot that, picked something up in one box and that's fixed. It's in a very specific orientation. It doesn't move. It picks it up. Moves it over. Puts it on the thing in a very particular location. You know, it screws in five bolts and you're done.
We've moved from that to something like an AMR (autonomous mobile robot) where you give the robot very general instructions like, I need you to take this thing from point A to B in your warehouse and the robot goes through this environment.
It's not fixtured. It can kind of deal with stuff happening. I can deal with someone walking in front of a box, falling on the ground that it has to go around. It has its flexibility and that flexibility has enabled a lot of new applications, new domains.
Other examples would be agriculture. Robots in agriculture weren't really a thing other than maybe like an assembly line. We have this robotic assembly line and you toss a bunch of apples on it. Maybe, if you're lucky, it can get rid of the bad apples. Now we're looking at robots that go out in the field and we're starting to see people trying to send robots out to pick the apples and put them in the box. Then another robot takes that box to somewhere else and they get washed and then sent out to market and stuff.
And so lots and lots more flexibility and a lot, lot, lot cheaper.
Wong: There are a lot of things that have been changing even in the past 10 or 20 years, but what do you think are the one or two most impactful things that have happened in robotics?
Scott: Impactful?
You have the increase in compute. Computers have gotten faster. They do more things in parallel and that increase in presentation and speed has enabled people to do much more complex things. We found all these new computer vision things that that enabled that.
When I started my career in early 2000, it was really, really difficult. Like the unsolved problem was, I have a room, I have a camera. I want to basically be able to build a 3D model of that room. That was not a thing that was possible in 1999.
We've gotten to the point now where you can go into a room with a camera around and it builds a 3D model. The robot can understand the space it's in.
Then with some of the deep-learning stuff it's gone to, I have the space. I know what the space looks like and I know what the things are in the space and I can find different things.
I think another big change has been the big use of simulation, which was around in some capacity for quite a while but robots are hard to work with. If you're a software developer, maybe you'd write a bunch code. Press compile and the thing works or doesn't. You write a bunch, you fix it, you do that over and over again. You get stuff done really quickly.
With robots, it's harder because you got to load the code on the robot. Run the robot for it to do the thing. Check that it did the thing, did it do the thing correctly, and then try again.
So when you use a simulator, run a virtual robot on your computer, you can close that loop a lot faster, but you can also, in theory, run ten different robots at the same time to see if it really works.
It's kind of like an imagination for robots, like they can see it in their head. You can see if something's going to work before you actually do it.
So I think those sort of dual things like more advanced perception to figure out rooms, figure out things, being able to map a room and then the ability to sort of have a mind's eye to figure out where and what you're going to do before you're going to do it and whether it's going to work.
Those are probably the biggest.
There's some hardware innovation and stuff, too. Everything's just gotten better and cheaper.
Wong: Okay. So what are you doing now at Open Robotics? And what did you do when you started?
Scott: I do a lot of the sort of community outreach work. I keep the websites going. I keep our forums going. That sort of thing.
Open Robotics, broadly speaking at this point, has not changed that much. We're pretty consistent.
I was actually just talking to somebody about this. It's kind of diversified. So we do everything from these big grand robotics competitions we helped put together. We built the simulations for them. We built the like infrastructure to conduct the simulator or conduct the competitions. We do other stuff, too.
We make ROS, the Robot Operating System, and Gazebo. People on occasion come to us and say, I'd really like this feature or I'd really like you to help us build this part of the robot.
You're the expert, you built it, you're the expert in it. Can we get your help?
We do smaller contracts sort of things where cool startups come in and they say: Hey, we're working on this thing. You guys are kind of experts. You've seen a lot of these. Can you do like a design review for us?
So we'll show you the robot. We'll let you guys poke around the code and you give us recommendations on how we can improve what we're doing. That was a really fun.
We have other clients. We do some robot installations here and there. I can't necessarily talk about all of them, but we do larger sort of robot deployments. We've actually been working on this system lately.
So one of the things that's changed fairly recently is it used to be, other than a factory, you'd have one or two robots doing very boring things.
For example, I have a robot mop and a robot vacuum. A lot of other organizations, like a hospital, might have a cleaning robot, a sterilization robot, a gofer robot, and food delivery robot. So you have all these different robots, all these different fleets, like a bunch of the same kind and they all have to work together.
They come from different vendors. They work in different ways and people needed some orchestration software to basically keep the robots from like running into each other. It makes it easy to deploy them. It makes it easy for them to work in the environment.
Say you're in a hospital and there's an elevator. The robot needs to be able to go up and down the elevator. Well, how do you make that happen?
It was like kind of an unsolved problem and another one is you have these eight different kinds of robots and you put them in a hallway and they just do whatever they want, right?
They're going to clog up the hallway. People are going to get mad. So you have to tell them to go on this side of the street only. You yield to this robot and this robot yields to this robot and they all yield to the people and that sort of thing. Solving all those like problems that you encounter once you have a bunch of different robots.
Wong: Excellent. Well, you mentioned ROS and Gazebo before. Could you tell us what those are?
Scott: Okay. So ROS stands for Robot Operating System.
Which is it? It's a misnomer because it's not an operating system, but it is also kind of an operating system.
Really, at the end of the day, what it is is a collection of all of the algorithms and infrastructure and tools that you need to build a robot.
So you have all these different parts in a robot. You have the cameras and the LIDARs and if you have the mobile base with wheels on the ground. You might have an arm somewhere that has to move around.
It's the glue code that has to move all the communication between them. There's primitives for building basically a program that ties all that together. There's primitives for building those primitive or just primitives for building that integration between the components. There is visualization tools and logging tools. People build packages that make it easy to deploy machine-learning models on your robot.
It's all of those sorts of bits and pieces. So it's a toolbox, really. You take all the tools and bolt them all together and you build a robot.
Gazebo is a simulator. Really, what it is at the end of the day, it's like a video game that's very, very tailored to a robot and when I say a video game, I mean that you can visualize things with physics and then you can put a fake robot in that video game and have it run robot software in a realistic way and see what happens.
We have simulations of warehouse environments where the robots are going around moving boxes and stuff and so you pick up the boxes, they have weight, they have like weight distributed in different spots.
You have to pick them up in a very specific way and the lights might go out or the floor might get slippery in spots. You can kind of simulate all these different parts and see how the robot behaves.
Wong: So what are the hot research areas now?
Scott: The hot research areas? There's always that the deep-learning side of things because, like I said, the perception side of things has been pushing the envelope for quite a while.
So initially it was, can we build a map? How can we build a map? What does that map look like? Can we recognize things? How many things can we recognize? How well can we recognize them?
There's a lot of work now on making that process easier and more robust. There's crazy, crazy stuff going on right now. It's like I take a picture of something and then I can basically build a 3D reconstruction of it which is really an important problem to solve. So that's important.
In robotics there's the so-called the sim-to-real problem. You have the simulation, you have the real world, and they don't match perfectly. Like the world's a complex simulation. It's kind of simple and how do we build stuff in simulation such that we know that when it works in the simulation it's going to work perfectly in the real world.
That set of problems, like making that whole thing run really easily, is important. Otherwise in robotics, there's also just the field of control generally.
So when I was an undergrad. I actually worked in legged robotics. It wasn't quite the Boston Dynamic robot, but very similar, sort of like six-legged robot. That was very, very researchy.
You just trying to get it to work and now you're starting to see people build real ones.
Oh, it's downstairs. So I have toy one sitting around here. So that's all control theory and so making those sorts of motions happen, particularly in stuff like picking stuff up, we sort of neglect how difficult it really is to pick things up.
Humans are really good at it and it's one of those things that is really hard to teach a machine to pick something up.
So those are fields that are actively being researched right now.
And what's cool about ROS in general is that it's cool that we can see that a student, like a grad student, build something and then somebody will sort of take that initial kernel of working code and iterate on it and fix it up and clean it up and work on it, and work on it, and work on it.
It provides this sort of pipeline to move that thing from the academic world where it's not necessarily perfect. It just kind of works to getting it all the way up to the point where it works for something like a business and that's the really cool part of our job is that we get to see that whole process of going from research and development, like academic stuff, to putting stuff out there in the real world at scale, and it's really satisfying to see that whole process happen.
Wong: Excellent. Well, thank you for filling us in and your work in Open Robotics and what's been going on for the past few years. It was really interesting.
Scott: You're welcome.
