For millions, if not billions, of years, life found a way. Through evolution, natural selection, and cosmic random chance, as some theories purport, simple organic compounds morphed into single-celled organisms which morphed into multi-cellular life, and so on into the sprawling world in which we (and plants, and animals, and bacteria, and viruses) live today.

But Piper Welch — a PhD student in UVM’s College of Engineering and Mathematical Sciences — is looking to understand, and eventually design, the evolutionary course that organisms can take. She’s currently working with Josh Bongard’s team of computer scientists who are designing and building “Xenobots,” the world’s first self-replication living robots.

Welch, who was recently awarded a National Science Foundation (NSF) Graduate Research Fellowship, hasn’t always been interested in designing Xenobots with supercomputers. Like most of evolution, it happened in small steps, starting at Carleton College in Northfield, Minnesota, where she completed her undergraduate studies.

“During my undergraduate years, I joined a laboratory that works with evolutionary techniques to evolve different populations,” Welch said. “My interest in biology and evolution transferred to an interest in evolutionary computation.”

Welch finished the first year of her PhD in computer science in May, working with Bongard and other researchers at Tufts University and UVM’s Vermont Advanced Computing Core (VACC) to design swarms of xenobots to perform specific tasks.

“We create random populations of beings and evaluate them for whatever fitness metric we’re using, which could be locomotion, to move in any direction, or move in a circle, or go up a ramp,” Welch said. “The top performers remain in the population, and the ones that perform poorly get deleted. Then we use the principles of sexual reproduction to blend the population and possibly find a better one. We mutate them and let them reproduce and then reevaluate that new generation.”

Like most scientific processes, designing and evolving xenobots isn’t a quick endeavor. It takes time, energy, and resources. It’s an even more arduous task when there isn’t a clearly defined stopping point for a xenobot’s evolution.

“That process keeps going until you find the optimal solution or until you’ve decided it’s gone on for long enough. In our case, we run it for however long because we don’t know if there is an optimal solution,” Welch said. “But just like in natural evolution, it’s not guaranteed to find the optimal solution. Our research is really based off of the fact that there is no intuitive solution to these problems.”

The ”problems” to which Welch is referring is programming the xenobots to perform a specific task. The problems compound when the xenobots have a limited range of mobility.

“We know that bots have three motion types: a circle, a line, or an arc. So we're kind of looking at how do we put these different three types of bio bots together, to have desired behavior outcomes and then we can control them on the swarm level,” Welch said. “ We rely on the evolutionary methods just to tell us what is a good solution to this when humans don't have an intuition about what it might be.”

Another problem is the building of the xenobots themselves, which Welch says is a hands-on process at the surgical level.

“Previously, we have had to design individual bot morphologies with artificial intelligence. Then, a microsurgeon at Tufts would have to impress our designs on each individual bot,” Welch said. “For example, if we wanted bots to collect particles in an environment, we give each of them a Pac-Man-esque body design. The process of sculpting their bodies is very time consuming, and hinders the scalability of our work. What I hope to discover is a way of controlling the behavior of biobots without having to do this — by selecting for behaviors at the swarm level.”

Welch, and the research team of which she’s an integral part, use VACC’s computing power to simulate and evolve the xenobot swarms.

“We use physics engines to simulate the bodies. Each individual bot has a genotype. Each genotype is simulated in the physics engine, where it’s evaluated on its behavior. But the evolutionary algorithm itself, they can be really simple. As long as you have your fitness function, you know what your evaluating them on,” Welch said. “It’s basic computer science, just loop over them and say, ‘if this organism is higher than this, then keep it and if not, delete it.’”

According to Welch, there are several ways living, self-replicating robots could be applied to tackle seemingly impossible challenges facing the modern world like pollution, oil spills, and medicine. According to some estimates, there is an estimated 75-199 million tons of plastic waste in the ocean, with over 16 million tons being added every year. What if xenobots could collect it?

“If we could use the bots to collect microplastics into a pile until the point that plastics are buoyant, they could float up to the surface and you could use some type of more traditional bot to come and collect the larger microplastic particles,” Welch said. Welch also described how xenobots could be used to detect oil spills.

“It's an area of research in which botanists are looking at putting different plants along oil pipelines,” Welch said. “Traditionally, people have to patrol up and down oil pipelines. But if we were able to put the xenobots in an environment such that they reacted to some input like oil in that environment, they could release some gas that had some effect on the the flora around and you could tell from aerial imagery where the leak is.”

Welch and the research team has discovered that the xenobots, as miniscule as they are, can carry a non-insignificant amount of weight, like medicine.

“We've discovered bots that can carry payloads, so they could be used for intelligent drug delivery,” Welch said. “If you need chemo at a particular site, we could possibly engineer the bots to seek that site and deliver the payload. The same is true for pain medication for a particular site.

Welch’s proposal to the NSF centered around controlling the xenobots, highlighting the difficulty in controlling them.

“I proposed a technique of controlling the xenobot behaviors through careful swarm compositions,” Welch said. “If we cannot impress designs onto the individual bots based off of manual shaping of the bots’ bodies, can we produce swarms of bots that might have desired behavioral outcomes based off of careful swarm compositions of the bots?”

Before Welch can answer that question when her research starts again in September, she’ll be presenting her research at the International Society for Artificial Life’s ALIFE Conference at Hakkaido University in Japan.

“For the rest of my PhD studies, I’ll be working and wondering what can these xenobots do?” Welch said. “Is there a limit to what they can do?