AI-Designed, Living Robots Can Self-Replicate

AI-Designed, Living Robots Can Self-Replicate

AI-Designed, Living Robots Can Self-Replicate 768 432 IEEE Pulse
Author(s): Ying Fang, Greg Orekhov, Zachary F. Lerner

In 2020, a research group made the stunning announcement that it had built programmable organisms—living robots they called xenobots—out of biological cells, and these xenobots could work together to perform simple tasks. Now, less than two years later, the same group has an even more astonishing announcement: The xenobots can autonomously self-replicate in a way that is completely different from any other living organisms, and their progeny are functional and similarly able to self-replicate [1].

“What we are showing here is not something that any natural system on Earth does: It’s not producing progeny directly, but rather it’s collecting material and building copies of itself. We don’t see any organisms on Earth that do this, so it was a huge shift for me to think that this is an entirely different way of replication,” said group member Douglas J. Blackiston, Ph.D., developmental biologist and senior scientist at the Allen Discovery Center, Tufts University, and the Wyss Institute, Harvard University.

“This is a discovery of a new form of replication, but this form of replication is particularly amenable to artificial intelligence (AI) control and design,” said group member Sam Kriegman, Ph.D., postdoctoral researcher at the Allen Center and the Wyss Institute. He is the lead author of the Proceedings of the National Academy of Science paperthat details the group’s new findings.

From cells to robots

The researchers’ work on xenobots came to the fore in January 2020, when they reported the generation of the first xenobots [2]. That work began in the Evolutionary Robotics Laboratory of Joshua Bongard, Ph.D., professor of computer science at the University of Vermont (UVM), where Bongard and Kriegman were running computer simulations of virtual organisms on the Deep Green supercomputer cluster at UVM’s Vermont Advanced Computing Core. Through the simulations, they hoped to explore the evolution of certain movements and development of rudimentary tasks.

Blackiston saw videos of Bongard and Kriegman’s simulations, which complemented his own ongoing work to understand how cells and tissues develop and form collectives. “These were simulations of soft-bodied virtual creatures that, to me, looked like something we could build out of cells, so my question became: Could we build something that the computer designed, and do it from the ground up using developmental biology techniques?” Blackiston recalled. “The answer at that point was ‘That’s impossible. You can’t build an organism from scratch.’” But he did … and quickly. Within about two weeks of first asking the question, he collected some frog cells, employed tiny forceps and other mechanical and chemical cues to nudge the cells into the algorithm-determined design, and then let the cells take over to complete construction of the first xenobots (Figures 1 and 2).

Figure 1. Using an AI algorithm and the Deep Green supercomputer cluster at the UVM’s Vermont Advanced Computing Core, researchers ran simulations of virtual organisms to find the best shapes and sizes for performing certain simple tasks. Here, the AI proposed an elbow-macaroni shape (pink) to pile materials, in this case simulated adhesive stem cells (green). (Image courtesy of Douglas Blackiston and Sam Kriegman.)
Figure 2. For scale, the AI-designed living organisms—xenobots—are shown in a petri dish over a dollar bill. (Image courtesy of Douglas Blackiston and Sam Kriegman.)

“The great part about this project is we are not genetically programming everything in a very reductionist way,” he remarked. “Rather, we’re using a combination of some developmental biology tricks to help shape and put the cells in the right place.” At that point, self-assembly takes over as the cells and then the tissues “do the connecting, building, and behaving on their own,” he explained.

In that January 2020 paper, Blackiston, Bongard, Kriegman, and Michael Levin, Ph.D., director of the Center for Regenerative and Developmental Biology at Tufts (Figure 3), showed they could build computer-designed organisms. The resulting xenobots were small blobs of cells that could scoot around the bottom of a liquid-filled dish on twitching “legs” that were “driven by the cardiac tissue pulsing, creating volumetric actuation, and pushing off the floor of the petri dish,” Kriegman described. The xenobots could also shove tiny synthetic particles here and there. In addition, the researchers demonstrated that the xenobots’ behaviors matched the simulated predictions.

Figure 3. Xenobot research group includes (left to right) Joshua Bongard, Ph.D., professor of computer science at the UVM; Michael Levin, Ph.D., director of the Center for Regenerative and Developmental Biology, Tufts University; Douglas J. Blackiston, Ph.D., developmental biologist and senior scientist at the Allen Discovery Center, Tufts University, and the Wyss Institute, Harvard University; and Sam Kriegman, Ph.D., postdoctoral researcher at the Allen Center and the Wyss Institute. (Photo courtesy of the Institute for Computationally Designed Organisms.)

The next milestone paper came in March 2021, when the researchers expanded on the xenobot forms and behaviors, focusing especially on upgrading the xenobots from walking to swimming by adding cilia. “In this case, we went from biology to simulation,” Kriegman said. Instead of trying to model the shape or alignment of the cilia to bring about a particular directional movement, for instance, they focused on predicting how the xenobots would interact with one another and with their environment. “And we found that if we just simulated the movement as random, but we put certain-shaped xenobots together in a dish and let them interact with even smaller synthetic particles in those dishes, we couldn’t predict exactly where each individual xenobot was moving, but we could predict, on average, the size of the piles of particles.” In other words, the researchers could build a living organism to perform a very specific task, which in this case, was to make piles of a certain size (Figure 4).

Figure 4. Here, the xenobots (C-shaped; beige) push loose stem cells (white) into piles as they move through their environment. (Image courtesy of Douglas Blackiston and Sam Kriegman.)

Self-replicating xenobots

The group’s latest finding is an extension of the work reported in March 2021, but the xenobots are now piling up stem cells instead of synthetic particles, and if left alone for a few days, those stem-cell piles become xenobots themselves (Figure 5).

“Part of the idea with this experiment is that there are various inputs to the system, including the number of stem cells and the number of xenobots, and those numbers can be optimized to make self-replication better or worse,” described Blackiston. “Under the conditions we started with, which were artificially set by us, we would typically get one or two generations of self-replication, so the parents would make offspring that were a little bit smaller, and those offspring would make slightly smaller offspring, but that was only because of the conditions that we set. What the AI did was use those exact same starting conditions, but create shapes and designs that allowed better piling behavior and more rounds of self-replication to happen, so we would get to four or five successive generations.”

Figure 5. Stem-cell piles can become xenobots themselves. Here, a parent xenobot (C-shaped) is beside an incipient offspring, which if left along for a few days, will become a replication-capable parent. (Image courtesy of Douglas Blackiston and Sam Kriegman.)

The idea of self-replicating machines has been around for decades, noted Bongard. “John von Neumann proposed them back in the 1940s, when he came up with this thought experiment where you would seed the solar system and beyond with self-replicating machines, but the assumption for von Neumann—and pretty much everyone else—was that the machines were going to be metal, plastic, ceramics, and circuitry. What we are showing in this article is, surprise!, it may be possible to make von Neumann machines and it might be easier than we thought, but they’re not going to be made from metal, plastic, and ceramics.”

Previous research groups have produced prototypes of machines that could build copies of themselves, “but they’ve all been manually designed and they’ve all been made out of electromechanical parts,” Kriegman added. “We’re showing this robots-building-robots form of self-replication, but out of biological components.”

For Levin, the latest work has several important aspects. “First, there’s this very surprising discovery on the biological side, and a deepening understanding of how plastic living systems are—how cells placed into new environments, despite a normal genome, can find new ways to solve problems and become a coherent, behaving proto-organism. And then once that discovery’s been made, it was almost immediately turned over to an AI, which will help us find the control knobs to try and enrich or accelerate or generalize in living tissues,” he said. “I think it’s an interesting and unique mash-up of biology and AI in that way.”

It also raises deeper musings about life, Levin noted. “Philosopher William James advanced this definition of intelligence as the ability to reach the same goal by entirely different means, given perturbations that have happened around you. With xenobots, we’ve made sure that they do not have the opportunity to replicate the way that frogs normally replicate, and it turns out that they have found an entirely different way of doing it,” he said. “So, while nobody’s claiming that the xenobots sat around agonizing over how are they were going to make copies of themselves, it turns out that they can accomplish the same outcome but through a completely different mode, given changes that have happened to them. And so, I think the fact that these things have been deprived of the ability to replicate the way that frogs normally replicate, and yet the system has found a way to do it, this is a great example of an unconventional kind of minimal intelligence: plasticity and problem-solving.”

Blackiston is similarly introspective. “Right now, I think we have to go back and start revising the definition of replication. What does it mean to reproduce? Sometimes I even forget at this point that these are genetically unmodified frog cells, and not some strange organism we dug up from the bottom of the ocean. These are cells from a model laboratory organism, but the cells have been leveraged in a certain way to allow this new behavior to emerge. It is very, very amazing.”

Future of xenobots

With the combined findings of the past two years, the four researchers are looking ahead to expanding work on xenobots.

Bongard is interested in the role of xenobots in shaping the overall future of AI and robotics, notably nonbiologic AI and robotics. “There’s a growing trend in robotics and AI to focus on how good they are, and conveniently forget how bad AI and robots can be in the worst case,” he said. For instance, when an autonomous car encounters a surprising or novel circumstance, it may not respond well and could have dangerous and even lethal consequences. Alternatively, life is very good at responding to the unexpected and unknown, he contended. “There’s the tongue-in-cheek phrase that ‘life finds a way,’ which for some people is kind of a scary thing, but is actually a very positive thing, and it’s completely missing from most of our robotic and AI technology. If we as a society want safe technology, those are the metrics and the regulations we should be focusing on: How well does a machine do under the worst circumstances, and I think xenobots in particular, and biology in general, have a lot to teach us about that.”

Kriegman and Blackiston are pursuing work to explore the xenobots’ ability to sense their environment, which would influence their behavior. Kriegman is looking into using sensing to design and control xenobots, and also making xenobots sufficiently intelligent that they are able to do useful work without direct intervention. Blackiston hopes to develop xenobots for specific tasks, such as pollution control. For instance, he said, a xenobot that can sense a particular water contaminant could potentially respond by immediately releasing a pollution-mitigating substance. “It’s ‘to-be-determined’ on exactly how we can take a more applied approach to engineering new sensors and the best way to go forward, but that’s the next step for me.”

Levin has a broad scope. “What I’m really interested in is getting as far as possible away from the default outcome. I’m talking about making xenobots out of new cell sources, such as cells from different species that have never lived together before, and with novel inorganic components, perhaps smart materials or nanomaterials, to make configurations that have never existed in the evolutionary stream on earth,” he said. “That will let us study their cognitive capacity, whether they have memory or preferences, and how we can learn to program a cellular collective toward an outcome, where there is no history of telling these cells how to work together, so we can basically manage this incredible novelty for the purposes of regenerative medicine, other medical uses, and of course for synthetic kinds of applications.”

Promise, perils, and philosophy

Throughout all this exciting work, the researchers have kept in mind that they are dealing with things that are alive, Blackiston assured. “Whatever the cell type, whatever the organism, we must act as moral agents in the design and deployment of any technology involving living cells or structures, so we certainly collaborate with ethicists, we have review boards that are made up not just of scientists, but other members of the community who have provided oversight before we even started our experiments.”

Nonetheless, the researchers get their share of questions about whether xenobots could be a threat to humanity by reproducing out of control, or evolving potentially dangerous new functions. “On the human safety of xenobots, I fall strictly on the side of this being nothing to worry about,” Blackiston said. “This technology is incredibly benign. For instance, there are more cell types in a tiny skin biopsy than in any of our xenobots, and you’re not worried about your skin biopsy taking over the world. We’re at a level that is so far down on the scale of what to be worried about that there’s no conceivable risk.” He added, “As the technology exists right now, it’s quite a benign, nongenetically modified technology that we’re using.”

Levin agreed that xenobots pose no threat, and said the focus should be on their great potential. “This is really an opportunity to reduce human and animal suffering by gaining the kind of information that I believe is going to be pivotal for regenerative medicine, as well as other medical and synthetic applications, going forward. To me, that is a critical part of any kind of a risk assessment for new technologies: To remember to account for the good they can to do improve the status quo.”

Xenobots will also engender a great appreciation for and understanding of life, Bongard added. “We feel like we have a good handle on decision-making, agency, consciousness, and free will, but as we make these more and more intelligent machines, we become increasingly uncertain about what those terms mean, and whether those behaviors are unique to humans. I think xenobots already are pressing on that sore spot even more heavily now, because they live in this world between something that’s recognizable—cells from an organism, a frog, that have evolved—and something ‘other,’ as the name implies,” he said. “It’s increasingly challenging our understanding of what it means to be human, to be an organism, to make decisions, and to love, hate, and feel. It’s very philosophical, but this is one of the things I love about robotics and AI, and I think xenobots will complicate this discussion moving forward and make it much more interesting in the years to come.”

Complicated and interesting are good descriptors of the future, Levin agreed. Until now, it has been easy to delineate life from nonliving machines, but emerging advances in bioengineering, chimeric technologies, and synthetic morphology will change that. “We are going to be surrounded by creatures that don’t look like anything we’ve seen before, and that don’t have a specific position on the phylogenetic tree,” he asserted. “Xenobots are just a first herald of that.” <

References

  1. S. Kriegman et al., “Kinematic self-replication in reconfigurable organisms,” Proc. Nat. Acad. Sci. USA, vol. 118, no. 49, Dec. 2021, Art. no. e2112672118.
  2. S. Kriegman et al., “A scalable pipeline for designing reconfigurable organisms,” Proc. Nat. Acad. Sci. USA, vol. 117, no. 4, pp. 1853–1859, Jan. 2020.
  3. D. Blackiston et al., “A cellular platform for the development of synthetic living machines,” Sci. Robot., vol. 6, no. 52, p. eabf1571, Jan. 2021.

Access the Full Paper on IEEE Xplore®

Sign-in or become an IEEE member to discover the full contents of the paper.