Today’s factories are largely single-task operations designed to produce one thing at massive scale. Automated assembly lines are set up to churn out the same products day after day, and industrial robots are tediously programmed for specific, repeatable tasks. Reprogramming those robots to take on different duties can take months or even years, which makes changing a factory’s setup often an impossibility.
But given consumers’ increasing demand for mass customization, the ability to reconfigure a factory’s setup will soon be that much more necessary for manufacturers. Fortunately, advancements in machine learning and robotics are revealing some interesting ways that industrial robots, and thus factories, might just be able to reprogram and reconfigure themselves. Watch the video to learn about some of this exciting research into machine learning and robotics.
Mike Haley, Head of Machine Intelligence, Autodesk AI Lab: Brickbot is a project that we began about two years ago to explore whether we could teach a robot how to build with LEGO bricks. The idea was, if you imagine a robot able to learn to assemble anything in the world, in the same way that a child assembles LEGO, then we could literally redefine how robots work in any industrial setting. Today, if you visit a factory, there are teams of people that spend months and years programming industrial robots to do just one task; it’s incredibly tedious, unbelievably complicated, and very error prone.
Yotto Koga, Software Architect, Autodesk AI Lab: The pitch that we’re trying to take with Brickbot is to use machine learning to take sensor data and then infer what’s going on in the environment to then have the robots act accordingly and act adaptively. There are a lot of things that you do when you’re working with robots: It’s some programming; it’s actually mechanical design, as well, making things, fabricating pieces that the robots may need to be able to do their task.
We’re looking at ways to make robots easier to use so that we can put these sort of assembly lines together and make them accessible to more people—not just these big companies that have deep resources to do this kind of thing. Traditionally today, if you look at, say, a robotic assembly line, something that’s putting together a car, a lot of effort is put into engineering that line to be what’s called very deterministic. Everything has to kind of be in its place for that system to work. If you change the design or you change the parts that go into that design, you have to reengineer everything so that those new parts are accommodated and made deterministic, as well.
Haley: We design factories today to effectively be single purpose. And the factories of tomorrow are not going to be single purpose. The factories of tomorrow are going to adapt to the needs at any one time. You might decide overnight to redesign your product. And you know what? By the next morning, the factory’s learned how to deal with that design change, and it’s ready to go.
The realization with Brickbot was that we could actually train industrial robots in a purely digital setting incredibly fast because it’s digital. It doesn’t have to be in real-world time. Because it can do it millions of times faster, you can train the robots more robustly. They can learn better; they can learn faster. So now you apply that learning not necessarily to LEGO, but to any industrial environment where maybe you’re assembling a car or a piece of an airplane or electronic device, whatever. The same thing follows. You’ve got a 3D model of those things; the robot has to interact with it; it can learn what to do.
We’ve now reached a point where certainly some of the stages of what we’re doing in Brickbot are fairly robust. So we’re now beginning to test them in industrial settings.
Koga: We want to move beyond just Lego bricks. So we want to start looking at more realistic scenarios. So there’s all kinds of different brick types, like wheels and gears, windows, rooftops, and so on.
Haley: Designing buildings, designing products are very, very complicated things. But you have to be a specialist to do those things, and as computers get more powerful, computers should be able to take on some of that specialization and make these things more accessible to more people.