HOD LIPSON’S ROBOTS ARE EVOLVING

If I could purchase stock on human beings, I would close my savings account, cash in the savings bonds that my grandparents have given me since I was an infant, take out a very large loan, and use all of it to invest in Hod Lipson. He’s the Director of the Computational Synthesis Laboratory at Cornell University, and, like most directors of Ivy League research facilities, he works on stuff that only about .003 percent of the world can wrap their heads around, but will probably affect everyone.

Among other topics, he studies evolutionary robotics, which is a subset of Artificial Intelligence that focuses on developing robots that independently learn design optimization through a Darwinistic war among the many algorithms inside of the robots’ brains. Lipson represents a broad field that has the potential to truly change the way human beings interact with machines–just like all of your favorite science fiction novels predicted. Much of the research towards evolutionary robotics is centered around practical innovation, suggesting that human beings are limited by the power of our design. On a much more philosophical level, however, evolutionary robotics may allow robots to develop complex social traits that are impossible to program, and may actually magnify a few newly discovered laws of nature.

Someday, he very well may be known as the father of modernity, or, if the project takes a morbid turn, the architect of the robot army that enslaves the human race. In the future, when Hod is frozen in a preservation chamber, robots will sing hymns on Sundays about Hod Lipson, vote for racist conservative politicians based on their uncompromising allegiance to Hod Lipson, and spend centuries of time waging wars because entire nations of robots won’t be able to agree with other nations on what to call Hod Lipson. In which case, his stocks would go through the roof, and I’ll be spending my time bathing in liquefied diamonds and purchasing snow leopard farms. This may have been the most important conversation I will ever have in my life.

Vice: I find everything behind evolutionary robotics incredibly intimidating. Can you explain it?
Hod: So, first of all, it’s a whole field. There’s more people and more groups like us. Basically, it’s the idea of using what we think of as biological evolution to design both the body and the brain, and to control the methodology of robots. Like many other fields of AI, there are two sides to the study. One is a very practical side, where we just want to design better machines with the hope that by doing this kind of biological evolution inspired optimization, we can design robots that people can’t design manually. The other side is much more philosophical, and is based on the idea that maybe this (evolution) is the way to design the intelligent robots that we’ve been imagining for 40 years but have never actually realized. Maybe it’s an alternative way of designing these very complex behaviors in machines that we can’t quite create manually. That’s the longer-term vision and a much more philosophical goal that researchers in this field are pursuing.

Are you personally focusing more on the philosophical, sci-fi side of evolutionary robotics, or are you working on an amalgamation of both sides?
Both, absolutely. In some of the projects, our entire goal is to have robots that can find a path faster than they would if they were programmed by a human design.

See, I find the robotic industry really fascinating because there isn’t much of a consumer side yet, but it seems to be something that you’re building into. I know there are massive projects like Honda’s Asimo, but it doesn’t sound like what you’re doing is anything like Asimo.
Well, there are a lot of robots all around us from Honda, Toyota, and other companies like that but they’re programmed you know? They’re machines. They’re not alive. There’s something fundamentally missing about them. Evolution is interesting because it can design more sophisticated things in the long run than what we could ever do manually. But also because it’s hands off, you don’t exactly know what you get, which is part of the mystery of life. When you look at a Honda robot that’s been designed by a team of engineers, you know that every single piece and every little part of what it can do has been approved and designed by a human. There’s nothing mysterious about it whatsoever. It’s blatantly transparent. When you’re dealing with an evolved machine, you don’t know everything about it. There are some things that you might know, but it’s a little bit of a black box. You can observe it from the outside, but you don’t know entirely what’s going on inside. This is the part of robots that fascinates us.

So if that is the case, and you don’t exactly know what’s going on inside of the robot, how exactly does this work? Is it all built on random algorithms? It sounds like a super complicated and intricate process.
Actually it’s a lot simpler than designing a robot the traditional way. Let’s say if you are evolving the brain of the robot, all you have to do is give the robot thousands of different, competing programs. Some of them will achieve better consequences than others, and then these consequences will breed more. The original will be reinforced and then there will be more copies of the original, successful program.

So is there a built in survival of the fittest mechanism, or is it an implicit function of these robots?
No. That process is something that we built in. The idea is that it takes whatever works, and makes more copies of the successful processes. After a while you get lots of copies, and the robots, on average, improve. You use this simple principle to make them faster and more efficient. That is why this thing that we have going on right now is a very interesting field. It’s improving the brain of the robot and improving the body of the robot on its own. Also, it’s not just the robot. I think a lot of things such as stock trading robots are improved by evolved principles with thousands of programs to see who can be better.

Since there seems to be such a prominent side of this field that has to do with the advancement of the robots that we imagined in Isaac Asimov books, have you gotten to the point yet where you are talking with the people who design animatronics? That seems to me the next step in developing.
Hmmm…I personally haven’t. There are a lot of people working in robotics who work hard to make robots look good. These are the people who fine-tune the movements to the point that it looks like it’s playing a violin, or where a dinosaur can move exactly how a dinosaur would have moved. It’s very specific choreography, and although that’s a great accomplishment, it’s not real. We’re after a machine that can learn on its own. Most of our robots don’t walk really well, and they’re certainly not state of the art in terms of how they look. We could probably manually make a spider robot that would run faster, look better, and become more reliable, but that’s not the point. The point is that the robot can do it on its own. The hope then is that if the machine can do it on its own, then it will ultimately get further than what humans could design, since we ourselves have limits.

Of the traits that you throw into the giant catalogue of possibilities, have you noticed one that particularly takes dominance, just like normal evolution?
Well, people are working on different types of things, but we’re most concerned with locomotion. I can’t necessarily point to any specific traits, like how some might prefer to crawl instead of walk. I think these things are uninteresting in the big picture. However, there is an interesting thing that happens when you allow robots to create models of themselves. They can sometimes create wrong and inflated models of themselves. Changes like that lead to a lot of interesting questions about psychology, and how humans create models of themselves that are not necessarily accurate all the time. Sometimes the self-image is inflated, sometimes it’s deflated, and sometimes it’s wrong to the point that it causes a whole variety of problems.

So you’re actually starting to see social problems in your robots as well? I feel like you could have a really profound discovery on your hands.
Well, we’re not really at that level yet, but you can see that as robots become more complex in their behaviors and thinking patterns, and you let them learn instead of function by programming, you’ll start to see these problems that are parallel to psychological traits in humans. From these behavioral traits, I think we’ll eventually get an amazingly clever and rational robot.

You think so?
I think the more sophisticated we get, and as people build more on this stuff, we will be dealing with all kinds of issues that arise from the complexity of the robots, such as these inaccurate self-models, which cause wrong assumptions and other social problems.

See, that’s the most fascinating part of this whole field of study. You’re starting to identify things that may be problems in a societal worldview, and you’re seeing them hyperbolically illustrated in these robots.
We’re looking at these robots with the goal to be able to ask the questions that we want answered about humans–such as what the origin of schizophrenia is, and finding ways to understand communication patterns and problems. It’s very difficult to answer these questions about humans because human beings are so complex. We can look at robots as a platform to ask and answer questions in a different domain. It’s not exactly like humans obviously, but it’s much more transparent, and we can use it as a tool to understand.

Have you been able to warp these behavioral traits with an external environment as well? Where you can elicit specific social traits?
We haven’t quite gotten there yet, but that’s definitely something we want to do. There is also an entire field of social robotics which focuses on how robots interact with people, how people interact with robots, and how each of them views each other.

What about how robots interact with other robots?
That’s something we’re working on as we speak. We are absolutely looking at how robots can reconstruct a mental image of what other robots truly are. There are opportunities for mutual presentation.

As a whole, obviously this field is incredibly exciting, but it also seems incredibly terrifying. Should people be scared of the future of evolutionary robotics?
Being inside of it lets me know that we’re very far from anything to worry about. However, it’s a very powerful technology, so I am very cautious, to a degree. It’s a big reason why I feel it’s very important that we do this research, keep it open, and discuss it with the public, so that everyone’s aware of it and can share their opinions. No one should do this kind of thing behind closed doors.

Are people doing this behind closed doors?
I guess I wouldn’t know if they were, but as far as I’m aware there’s a huge community where everything’s open. Like I said though, all the robots that we’ve been able to produce with these techniques are currently far inferior to those that engineers are making and programming manually. We haven’t caught up to mainstream robotics in terms of capabilities. It’s still a question of whether this is a worthwhile pursuit.

BEN MAJOY