These Researchers Used AI to Design a Completely New 'Animal Robot'

"Xenobots" are living, swimming self-powered robots that measure less than a millimeter across. They are evolved by artificial intelligence and built out of frog stem cells—and they could open new medical frontiers.

TRANSCRIPT:

Sam Kriegman: They are a living swimming self-powered robot that is less than a millimeter across.

Doug Blackiston: So some people were really very afraid of the technology. It's something new they've never seen before. Other people were really excited about the potential to treat a number of different diseases in humans.

Mike Levin: I’ve heard this idea that this kind of work is so-called playing God, but I get phone calls every week from people with the most heartbreaking conditions, and all of them are waiting for new science to try to improve life for all.

Josh Bongard: Are they robots? Are they not? Are they organisms? Are they not? They are forcing us to blur previous distinctions which might allow us to see the world in a new way.

Kriegman: The xenobots are designed by computers, and we built them in the real world out of real living tissues.

Levin: We call these into bots for two reasons: the frog that these cells are made of is called Xenopus Laevis. That's where the xeno comes from. Bot means robot because this is a profoundly interesting biorobotics platform.

Bongard: There is no remote control hidden off to the side. They're made from genetically unmodified cells, but they've been created by AI. And they are definitely not a normal animal.

Ritu Raman: Only in the past few decades have insights from biology fed into engineering and vice versa. A lot of the first applications in the field were very focused on medicine so thinking about how do we put cells together to make replacement tissues or organs for medical applications. But then they also started getting more excited about, well, maybe we could build things that don't already exist in nature.

Blackiston: And if you take that to the extreme, that is where this biobot project really came from. Can you put cells together and build a synthetic organism from the ground up?

Levin: It's actually a new kind of engineering, which is a collaboration with your material. When you work with cells and tissues, you are working with a material that actually has its own agenda. It has preferences, it has the ability to solve certain kinds of problems. It doesn't behave the way that wood and metal might behave where you know exactly what it's going to do.

Bongard: This brand new idea of building robots out of living cells seemed really, really difficult.

Blackiston: Our collaborative research on this project began around 2017.

Kriegman: What the original plan was, was to get computer scientists and developmental biologists in the same room under some funding to work together to try to create AI systems and robots that could work out in the real world, drawing inspiration from biology, which is so adaptive. But there was nothing about building robots out of frog cells. At the time, we were designing virtual creatures inside of a computer, soft robots that you could cut off part of its body and it would deform to recover its functionality.

Blackiston: As we were doing the study, I'd been watching Dr. Bongard and Dr. Kriegman's simulated models. But the things that they were building in their simulator looked really similar to things that we could build with frog cells. So could we build some sort of facsimile biologically?

Kriegman: I wasn't really sure that that was going to be possible. I might have brushed Doug off and said, you know, sure, go try to do that.

Levin: When Doug first said that he could make the various shapes that Josh and Sam created, I had no doubt that he could, because these cells are really motivated to work together to build something. They can adjust to all kinds of changes, all kinds of experimental perturbations to really get their job done. For example, he created tadpoles with eyes on their tails and showed that these cells could see.

Blackiston: So that got me interested in this idea of modularity or pieces of an animal that can be moved around like Lego and reattached somewhere else. And if you take that to the extreme, it's taking the pieces of developing frog and building something entirely new from the ground up. And so when I said that this is possible, I think that they probably believed it was a joke. And really, that's to a scientist throwing down the gauntlet. So now I must build it, right? I'll show them that this technology exists.

Bongard: Unbeknownst to us, Doug spent the week very carefully putting together some frog tissue under the microscope to try and build one of Sam's creations.

Kriegman: Honestly, I wasn't really sure what I was looking at at first. If this was just kind of a party trick, but it was completely unexpected. And we started to realize that if we can actually copy the movement that we saw in the simulation then we would be able to design these systems in the simulator and build them a reality. And that was very exciting.

Bongard: Doug is fantastically talented, but it still takes him about 4 hours to create one millimeter sized xenobot. Turns out that AI can make billions and billions of candidate designs for hours or days or weeks on a supercomputer which is much more efficient at finding interesting, useful designs that Doug can build.

Kriegman: This kind of design process takes inspiration from Darwinian evolution and natural selection and applies it to robots. What happens is that we supply the computer with these building blocks. We have one block and it kind of contracts and expands. This is simulating heart tissue, and we have another block that is just passive frog cells. And initially, the computer randomly assembles the blocks into a bunch of random designs, and then we supply a behavioral goal. What should the system do?

Bongard: At the very first experiment, what we started with was, we want a millimeter size machine that walks along the bottom of a petri dish. Basically, the supercomputer deletes the poorly performing creatures, makes randomly modified copies of the survivors.

Kriegman: And over many generations, this population of robot designs gets better at the objective. You can evolve their shapes, their material properties, their control systems. And this allows us to go through billions of years of evolution all at once.

Bongard: But eventually we get back the handful of champions to see which, if any of them can be turned into reality.

Blackiston: Every week, Sam provides me with a model of shapes that perform a particular function, basically a blueprint, almost like if you're assembling furniture or Lego. So it all starts with a developing frog embryo that's 24 hours old. And at this stage, all of the cells are still stem cells. STEM cells can become anything like skin or an internal organ.

But we have an atlas that maps these different regions of the embryo. And so over the course of the next 24 hours, I very carefully, with microsurgery tools, harvest different parts of the embryo and collect them into different piles. So these cells are naturally sticky. So if you take loose stem cells and put them together into a pile, they adhere into a sphere over time.

And so if you're very tricky and very careful and have very good fine motor skills, you can now take the sphere that has all sorts of different tissues inside and sculpt it to reveal some form the way that you would sculpt away a piece of wood. I think that there's this misconception that all of biology is this hard, rigorous discipline and that there's no creativity or beauty or artistry.

It's incredibly amazing and relaxing to look down the eyepiece and to build something that no one has ever seen before. But it's also important to the science to generate shapes that wouldn't be produced naturally.

Bongard: And it took us another year or so to be able to defend that. There was a match between the evolutionary algorithms, dreams or designs and what we were getting in reality. The moment we had that early data, I got up from my desk, my hands were shaking and just absorbed the implications that this was possible.

Levin: When I saw these creatures for the first time, I was absolutely blown away. They have spontaneous motion, they gather together and have really complex interactions in a group. They can heal themselves. So if you cut one almost in half, it will zip back up and that's just that's just the beginning.

Blackiston: So the first designs that I built in the lab were cardiac driven. So these are heart cells. And through contractions these walked.

Kriegman: But eventually we switched to cilia.

Blackiston: So these are small hair-like structures on the outside of the body.

Kriegman: It produces movement that is much faster. And using cilia, it allows the robot to swim instead of crawl And Doug, trying to visualize the movement of the xenobots that he built, placed dye particles on the bottom of the dish And what we noticed was that they will tend to create piles, which led us to the idea that they could be designed to be better and better bulldozers.

Blackiston: And that's an interesting behavior that can be leveraged for all sorts of interesting mechanical work. But Mike Levin had the idea of what would happen if these particles were replaced with stem cells. So this is the material that the biobots they themselves are made of.

Kriegman: And if the xenobots made big enough piles of cells, those piles could themselves form into little children bulldozers. So unlike every other living system that we know of, the xenobots do not reproduce through growth. We give them additional building materials and like a robot in a stock room, they stick together as materials to build a copy of themselves.

Blackiston: So just by starting the first generation as a C-shaped the spheres that were created were much larger, which created larger offspring and allowed us to get the five to six generations.

Bongard: We don't want uncontrolled self-replication. That's a dangerous thing. But for a roboticist, the realization that xenobots could self-replicate is a huge deal because roboticists have been trying to create self-replicating machines for a very long time.Seems that when you do this with living materials, it's suddenly much easier

Blackiston: The thing that caught me by far the most off guard in this entire project was the public response.

Bongard: A lot of the global media saw this as a huge breakthrough, which was followed by a massive outpouring of opinion on social media. Fear, paranoia, excitement.

Blackiston: It just went wild. We were getting tweets and Instagram posts and news stories from countries all across the world.

Bongard: The xenobots might escape from the lab and replicate out of control.

Levin: Now, a lot of people will they will say, well, this isn't natural or so-called things that shouldn't be created. But we really have to understand the ethics of this, much like any other biological experiment is really, really critical to think about before you do anything.

Carolyn Neuhaus: When I first read articles about xenobots, my response as a bioethicist was that I find it very difficult to sort of view these as morally different than other tissues in a dish. I understand that they're different from cells that we normally culture because they've sort of been given a task to do. Any manipulation that happens doesn't yet affect how an animal interacts with the environment.

It should always be on our radar that people might do nefarious things with the knowledge produced or the technologies that are used in this research. But that's not what people are reacting to when the public reads that there's a robot made of cells. I don't think people are sort of wrong to think Oh, that reminds me of Frankenstein.

It's not surprising that people continue to return to that trope of scientists creating things in the lab that are not able to be controlled because it's one that constantly is in pop culture. It's in the sort of social imaginary that informs our our beliefs and our opinions. Again, it's not going to happen in this case, but I think the lack of clarity and certainty about what exactly has been created definitely fuels a negative reaction among people who read about it.

So there is not scientific consensus about the definition of organism, and there is not consensus about what is a robot. The ambiguity about the terms then yields a moral uncertainty and a moral ambiguity. What's going on here? What is it? It's then wrong for that to be the final word. It's one input into the calculus of what do we owe to the creature that we've created?

What are the kinds of safeguards and oversight we should have on the research? And where do we go from there? Is the conversation that needs to happen at this point.

Levin: I really think about the old story of Adam naming the animals in the Garden of Eden. It's the idea that we have to figure out what criteria we're going to use to establish our relationship with these other beings. And from that perspective. This is the first step towards a new science of really finding the essence of new types of synthetic organisms by robotics that are going to be very plentiful. And in the coming decades, are going to be all around us.

Raman: There are quite a few challenges that we still have to tackle in this field. I'm a professor of mechanical engineering at MIT, and in my lab we specifically look at how we can build machines that we call bio-hybrid because they're part biological and part made out of synthetic materials. The biological robots that we're building are powered by muscle tissue so that every time the muscle contracts, you could get something that looks like movement.

The actual muscle tissue that looks sort of like a pink jelly like rubber band was formed using these mouse skeletal muscle cells that we had genetically engineered so that when we shine blue light on them, the muscle would contract. What we're really thinking about right now is how do we take all these other cell types like motor neurons and then get the neurons to grow into the muscle and form functional connections that can turn portions of the muscle on or off We could potentially use that to programed a clump of stem cells into any sort of shape.

Levin: The answer to this one question: what kind of signals could we give these cells to get them to do something different is really the gateway to a profound improvement in human health. Birth defects, traumatic injury, cancer, degenerative disease. All of these things could be solved if we had the ability to tell cells what to build. Once we understand how this works, we can test this in xenobots by showing that we can make cells build whatever structures we want, and that will go directly towards regenerative medicine inside of bodies hunting down cancer cells to regenerative repair.

And let's be clear. I don't think that actually anybody, of course, made out of frog cells are going to be in human bodies. But as we move towards human medicine, the next step is going to be to make these out of mammalian and eventually out of a patient's own cells.

Raman: Making a fully biological robot that could go inside the body and do some functional task to restore somebody's health. Their quality of life is certainly within the realm of possibility. But at the end of the day, we want robots that can tackle a whole bunch of different problems.

Bongard: We're optimistic that we can create more and more powerful AI that dreams up more and more complex, useful, maybe general purpose, biobots.

Kriegman: So the obvious next step from a robotics perspective is to add sensors so they can move towards or away from a stimuli.

Blakiston: We would like to engineer the system to respond intelligently to environmental cues of all kinds. So you could imagine creating sentinel bio robots that sense contaminants in the environment, and then you retrieve them and you have a biological readout of what's going on in the local waterway.

Bongard: We also have a whole bunch of technological challenges ahead of us. What we'd like to do, for better or worse, is replace Doug with a bio fabrication facility that can pump out millions or possibly billions of biobots. That might allow us to scale up this technology. But I think their most important long term implication is that they are new scientific instruments. They're a new type of telescope or microscope.

Blakiston: And so we have a whole separate research program that's spinning off of this to really understand some of the basic rules that cells use to assemble, to treat human disease.

Levin: And then, of course, we're going to try to augment their cognitive capacities by asking, what does it take to make a collection of cells better at navigating problems in various spaces?

Kriegman: How do we move from just a ball of motors to a ball or another shape of sensors, memory and motors? And then maybe we can achieve something that can learn and exhibit higher level cognition.

Bongard: That is going to help us to understand the foundations of neurological intelligence.

Levin:Think about it. We were all single celled organisms once. We were bacteria before that. And now you and I feel like an integrated, centralized intelligence with our own memories and hopes and preferences. There was no time at which a magic lightning bolt said, Okay, before you were just chemistry and physics, but boom, now you're a cognitive creature. That doesn't happen.

We can't roll the tape of life back on Earth, but we can create xenobots to ask how do small, competent subunits work together to make a larger mind? That is perhaps the most profound puzzle facing us.

Blakiston: It does feel like sometimes there's some magic involved in the design process, and it can feel quite big as these questions begin to spiral into larger and larger research programs.

Levin: There are so many questions here that are really pushing beyond known boundaries as they are themselves crashing all of our definitions of organism, of animal, of robot. None of these terms are appropriate now in their original meanings. But we are most certainly at the edge of many unknowns here. It's very exciting.

Kriegman: AI can not just design parts of biological systems like proteins, but they can design whole new kinds of organisms that look and act completely differently from anything else that has existed on Earth.

Bongard: Mother Nature has been working here on the planet Earth for 3.5 billion years, and despite her immense creative capability, she has only explored a very, very, very small part of morpho-space, the space of all possible organisms It would be beautiful and it is beautiful and wondrous to open our eyes wider, to understand not just life as it is, but to understand life as it could be.