Command Line Heroes - Robot as Software

Episode Date: September 21, 2021

Building a physical robot isn’t cheap—even when it’s the final version. Designing a robot and testing it over and over again? That takes a lot of tries. And likely more than a few failures on the way to success. Luckily, simulation software is reducing the scrap heap—and bringing down the costs of building robots from the ground up. Kevin Knoedler shares how simulation software allows him to program and design robots from home. And even though he doesn’t have the budget or support of major research institutions like DARPA, his robots still end up winning major competitions. Evan Ackerman points out that winning those competitions takes a lot of skills. But amateurs have more ways than ever to get started with robotics. Louise Poubel explains how much time, energy, and money is saved with robot simulation software—and how it’s not just for the amateurs. And Dr. Timothy Chung reveals how competitions like the DARPA Subterranean Challenge encourage innovators to advance the field of robotics.If you want to read up on some of our research on robot simulation, you can check out all our bonus material over at redhat.com/commandlineheroes. And follow along with the episode transcript.      

A dust storm is ripping across the surface of Mars, headed toward a humble set of buildings. The first human settlement on another planet. Martian dust storms like this one are incredibly powerful. As it hits, it tears open a wall of the habitat, flips over the solar array, and bends back the communication antenna. Years of labor ruined. Unless emerging from the dust cloud is a humanoid robot called Valkyrie. NASA's only hope. With the right programming,

Valkyrie can patch that air leak, deploy a new solar panel, and realign the antenna. But this level of robotic finesse has never been attempted. Luckily, back on Earth, a stay-at-home dad named Kevin has everything under control. The disaster scenario you just heard was imagined by NASA back in 2015. They laid it all out as part of their Space Robotics Challenge. Then they invited anybody and everybody to try and solve it. How do you program Valkyrie to execute all those fixes? It sounds like the kind of challenge only huge teams of people can tackle.

But that stay-at-home dad I mentioned, Kevin Knaedler, he really figured it out. He won NASA's challenge. And Knaedler never had access to a Valkyrie robot. He figured everything out from his kitchen desk while taking care of two kids. He could do that because robotics has opened up. It's now a software-first field,

a field where anybody with a laptop and the right know-how can be a hero. I'm Saranya Bari, and this is Command Line Heroes, an original podcast from Red Hat. This season, we're exploring the difference between robot fiction and robot fact. What did we imagine our robots would be, and what did they turn into in reality? In the movies, robots are often designed and controlled by a secretive elite. We imagine that you need endless resources and powerful institutions to bring something like WALL-E and Johnny Five to life.

And that's not entirely wrong. But we've discovered that a leap forward in robotics doesn't always require expensive hardware and giant labs. It can happen in the free and open world of software. The real robot future is being built by open source heroes all around the world. Dad, I'm going to be late for school. Contributing from the middle of their everyday lives. Doing robotics contests while parenting is interesting.

Kevin Knaedler lives in Newberry Park, California. When he won NASA's Space Robotics Challenge, his kids were 9 and 11. When they were younger, he could only do robotics work when they were asleep. Now that they're in school, he's got the bandwidth to, you know, save a colony on Mars. It's really just one computer. I don't think I'd be allowed to keep it on the kitchen table. I have a desk set up in the kitchen. We actually have all of our computers set up in the kitchen. So when we're all home together, we're all in the same room, either going to school or working on things and such. In between helping with homework, Knaedler's able to take part in major robotics challenges because he can take advantage of new software that's revolutionized the field,

opening it up to people like him. It doesn't require a huge amount of equipment or investment to be able to do that do a simulation contest like that. For one thing, the internet is at his disposal. Knaver can do arcane troubleshooting, drawing on a planet of experts. But there's also Ross, the robot operating system, which we mentioned in episode one.

Ross gives him a ready-made set of tools, a suite of crucial middleware. It really is a common standard that a lot of robotics packages are built on. And so rather than having to integrate and find what you need, a lot of that is already integrated and working within Ross. Just as important as that toolkit, though, there's a powerful new simulation software. That NASA contest, for example, was a simulation contest, meaning nobody had to build a physical robot. You didn't even need to have access to one of NASA's Valkyries. You could solve the whole thing using simulation software from the desk in your kitchen.

You had to pick these tools up off the table with a humanoid robot, get them oriented correctly within the limitations of that robot's arm, and then use them to search the wall for leaks. All these granular details could be manipulated using an open-source simulation program called Gazebo. Knapler only needed an NVIDIA graphics card to run it. And once you're working in a simulation, just imagine how that frees you up to experiment.

When you're talking humanoid robotics, for example, every time you run the humanoid robot into something or fall over, you can do tens of thousands or hundreds of thousands of dollars of damage. In Gazebo, you can rack up those damage charges hundreds of thousands of dollars of damage. In Gazebo, you can rack up those damage charges and not actually have to pay that bill. You can get that learning without having to pay all the costs in terms of breaking things. Trial and error can suddenly take place without any cost at all. Using the Gazebo robotic simulation system as a tool makes it possible for people working from a lot of different locations to be able to do robotics work.

You might be wondering, how real are these simulations? If it works in simulation, do we know it works in real life? Or are we just playing video games? Well, Knaver had the opportunity to go to the New England Robotics Validation location, where he ran his award-winning code on an actual Valkyrie robot. And it actually ran. We got it running within the first day by the support of the staff there and the quality of the simulation from Gazebo. And so rather than taking weeks or months to bring up tasks like that on a human or a robot, it was done in a day. And so that's really one of the big powers of Gazebo.

Knaedler won $175,000 for his perfect run in the NASA challenge. Now, before I paint too rosy a picture, I should point out that Knaedler is a graduate of MIT. He's got skills. He's put in the work. And as much as robotics is opening up with tools like ROS and Gazebo, it's not all the way there. Evan Ackerman, a senior editor at the engineering magazine IEEE Spectrum, gave us the amateur's perspective. ROS is certainly accessible to anyone who's willing to put in the time and effort.

It's just that if you're doing it by yourself, it's not intuitive. I think that the general problem is that Ross is a community project and it's designed sort of within the community for the community. And that community is people who generally are already robotics experts. Ackerman points out that you can always work your way up to Ross, though. If you're actually a beginner and you just want to have a good starting point, there are all kinds of robotics companies and products who are happy to help you with that. You know, Lego, for example, has a really robust education program. That's Lego as in colorful building blocks. They have

a robotics line for the uninitiated. And Ackerman has another warning for those who want to get into robotics. Even while you capitalize on all that software, don't forget that hardware isn't obsolete. The real physical world still matters. And that real world is full of chaos. And we can't simulate that. Our physics simulators just aren't good enough. So no matter how well your robot does in a simulation, the simulator itself is just not a good representation of the real world. And in some cases, that doesn't matter. You can get things done in a practical way, even if your simulations aren't exact. But keep this in mind, the real world is its own best model. There is certainly a huge gap between simulation and actual hardware. Hardware,

I mean, it's a little bit of a cliche, but people say that hardware is hard, right? And no matter how much work you put into your simulation, there is still going to be a gap there. And the size of that gap really governs how much you can learn from the simulation, how useful it is to keep doing stuff in simulation before you say, okay, like we really need to get on a real robot before we can make more progress than we have. Hardware does matter. And that's why Ackerman has been so happy to discover something called the Turtlebot, a low cost physical robot that uses ROS and lets amateurs like Ackerman run their software in the real world. Even though it's just basically a Roomba with a laptop on it and a Kinect sensor,

that's what it used to be. Don't downplay the turtle bot. Even a glorified Roomba has a role to play. It moves, it thinks, it senses, and it's a great way for, you know, like a university course or even a high school course to say, look, we have this hardware now, and, you know, our students can work in simulation

and then actually test out what they've been simulating on the real robot and learn about that gap. Devices like the TurtleBot show you in a visceral way whether your slick algorithm is as slick as you thought. Students can learn more. Entrepreneurs can prototype their inventions. And every robotic simulation gets a chance to prove that they're not just strings of abstract code.

They're blueprints for a revolution. Hardware and robotics is almost like a fact checker, a way to prove that our software stands up to the endlessly complicated world that we actually live in. But that doesn't mean simulations aren't practical. Sometimes they're the most practical, the most effective approach you can take. If you're trying a new algorithm, you may just like break that robot. It can cost you millions of dollars if you make a mistake on your algorithms, right?

So you really don't want to be trying like new crazy things in a physical robot. Louise Poubelle is the technical lead for Ignition over at Open Robotics, where they help people simulate before they build. And she's been finding that simulation is not just for indie robot enthusiasts like Kevin Knapler. Every player in the field, no matter how large or how well funded, has come to rely on simulation software. I mean, think about how simulation levels up your work. If you're developing a robot that is very expensive and you have a large team, not everybody can have access to the robot at the same time. But virtual robots are free. You can duplicate them. Each developer can have their own copy of the robot running in their,

like multiple copies, right? Running on the cloud, running in their own computer. Makes sense. No group has endless funds. And the more ambitious your project, the more even a wealthy organization is going to want to capitalize on simulations. Imagine if you're developing algorithms for a fleet of quadcopters. Like imagine trying that in the real world. Every single time you have to start up like 50 different robots and one of them like doesn't do the right thing and breaks and then you have to go and fix it. Like in simulation, you just restart and you start from zero. It's so much more convenient. Now imagine you're trying to develop a helicopter that can fly on Mars.

When NASA created a flying robot called Ingenuity, the sidekick to their Perseverance rover, they needed it to fly in the Martian atmosphere, which is way thinner than the atmosphere on Earth. By testing their helicopter in simulations, NASA scientists could see how it would fly on the red planet. And it worked! Ingenuity made the first Martian helicopter flight on April 19th, 2021. But whether the robot's an Earthling or a Martian, they're going to encounter something that a simulation couldn't prepare them for. Things in the real world, they are not perfect. They have faults. You know, parts are not exactly the size that they are supposed to be.

One common thing that people always talk about is like how hard it is to make a robot with two wheels just go straight in the physical world. Because, you know, you would say give both wheels the same speed and the robot's just going to go straight. But, you know, the wheels are never, they never have the exact same size, the exact same friction. Pubelle says that gap between simulated reality and actual reality is in fact shrinking. So there will always be a step of fine tuning for the physical robot. That's often becoming shorter and shorter to do this fine tuning, especially with people doing things like domain randomization and machine learning, where in simulation, you don't test just one scenario. You can test like thousands of scenarios.

So say you're working on that two-wheeled robot she mentioned earlier. By simulating a hundred different levels of friction, you can quickly try those wheels out on a hundred different surfaces, making sure your algorithm applies to each one. And then when you go to the physical robot, there's very little that can surprise you because you've made your algorithm be so robust to all of these different situations. And remember, as with ROS, the Gazebo simulation software is open source. It's becoming more user-friendly and more expansive all the time. People could fix Gazebo. People of different backgrounds, of various different use cases,

could look at Gazebo and say, hey, we can tweak it like this and it's going to make it better for everybody. That's another win for big organizations. Because when crowds of amateurs come join the party, their breakthroughs, their best practices, their hundreds of tiny advances push everybody further. An open, software-focused robotics field isn't just cheaper and faster than a hardware-focused field would be. And it's not just handy for amateurs. It's actually more powerful across the board.

So, when robotics began opening up, it started powering up too. And to maximize the potential of platforms like Ross and Gazebo, we'll need to invite everybody to the party. The worst thing that could happen is that a small group of people, all similar to each other, develop robots and they ship these robots to everybody else in the world. And just impose this view and the needs of this small group of people onto everybody else.

I think it's really important to democratize it a little bit more. Whether you want to democratize the field or you just want robotics to advance as fast as possible, the answer is the same. Get more people involved. Open things up and share. So how do we invite more people to the robotic software revolution? We're very much interested in identifying where those brand new ideas might come into play. Timothy Chung is a program manager at DARPA.

DARPA is the Defense Advanced Research Projects Agency. These are the folks who brought you GPS and, you know, the internet. They know how to build new tech. And Chung, who focuses on robotics, feels that group challenges, where people from around the world compete, are one of his best tools for inspiring innovations. For example… The DARPA Subterranean Challenge is a worldwide international competition really keen on identifying and discovering technologies for robotic systems that operate in complex underground environments. DARPA's Subterranean Challenge is similar to the NASA challenge I mentioned at the top of this episode. A way for a huge organization to tap the brainpower of the crowd.

It's a kind of underground scavenger hunt they've been running since 2019. Teams deliver robots that fly, walk, or roll their way through a system of caves, hunting for objects planted by DARPA. They're testing robots in underground caves because a million bumps and cracks can make them stumble. And they want a robot that can stay the course. Nature has a way of keeping things unpredictable. They want an all-round athlete that can rise to random challenges. DARPA wants to develop robots that can navigate the world's most difficult settings. And yet, despite the very physical nature of this challenge,

simulation software is still crucial. It aids in the development of the physical robots, of course, but there's also a separate, entirely virtual competition run by DARPA. That one's called the Sub-T Tech Repo. The virtual competition's kind of like your fantasy football league, where I provide you with a thousand Sub-T credits. You get to go to the SubT tech repo and pick and choose the different robots that you'd like to add to your fantasy SubT team and design your algorithms to best suit those robots. Those virtual competitors

then modify sensors or bring in new open source models that they find. They tailor things as they see fit. We've seen that type of vibrant interchange, and that's been phenomenal for the systems teams to see their robots being used. They might learn a thing or two from the virtual competitors. Cheng has already seen advancements in code and design philosophy come out of these challenges. It's so much more than a scavenger hunt. So remember Kevin Knaver? The guy who won that NASA competition by designing a Martian rescue mission from his kitchen? Well, same guy went ahead and won DARPA's subterranean challenge in 2019. And some of the teams he competed against

were funded by DARPA. He was David against Goliath. I was quite proud of myself for being able to win a competition against those funded teams. Interestingly, of the top four, three of them were self-funded teams. Software has been an equalizer. Robotics was, for decades, a field where only rich organizations could make big advances. But software has begun to change all that. It's still often those with time and education who get to really dive deep. But the gates are beginning to open. More people than ever are discovering they can contribute.

The availability of open source software is a huge deal in terms of making things easier. It makes things possible to do that would not be otherwise possible. And over at DARPA, Timothy Chung doesn't worry that those DARPA-funded teams were beat. In fact, the way he sees it, everybody wins. And meanwhile, every time Knaver enters a competition, the whole community learns from him and the bar gets raised for next time. He continues to have to up his own game

as the community continues to benefit and grow and excel. And, you know, that's one of the advantages of competitions is we like to call it coopetition, where you can cooperate and collaborate with your fellow collaborators and competitors alike. And that just elevates everyone's game at the end of the day. The old fantasies we had about robots, all those movies where robots were designed in secret laboratories, they really got something wrong. Secrets aren't good for innovation.

And software like ROS and Gazebo mean that every roboticist from every background is invited to the field. New and unconventional solutions can come from anywhere. When I think about the power that open source software has on robotics, the way it democratizes the field and pushes things forward, I have to remind myself that one day, these robots show up in real life too. The DARPA subterranean Challenge matters because, for example, DARPA might create robots that help us rescue humans from a mining accident or a burning building. The great thing about our robot dreams

is that one day they become real. And next time on Command Line Heroes, we're exploring one of the most real moments in robot history. The moment that huge and physical robots arrived on the factory floor. These weren't just simulations or abstract pieces of software. Factory robots were taking up space and deciding how much space they should take up became a global debate. Make sure to catch episode three and every episode by following or subscribing now, wherever you get your podcasts. I'm Saranya Dbarik,

and this is Command Line Heroes, an original podcast from Red Hat. Till nextris, Chief Strategy Officer. I've been a Red Hatter for about 25 years. And before your episode starts, I want to talk a bit about AI. The hot topic right now is foundation models. And those are important, but at Red Hat, we see them as just a piece of the larger AI infrastructure. And here's what I mean by that. Enterprises are built of hundreds or even thousands of applications. It's not hard to imagine a future in which those applications are being served by hundreds or thousands of models. Without a common platform for your data

scientists and developers, without a way to simplify some really complex workflows as you train, tune, serve, and monitor models, it can get overwhelming pretty quickly. And that's why we've built Red Hat OpenShift AI, a platform where everyone is working together on the same page to build and deploy AI models and applications with transparency and control. Find out how at redhat.com.