In January 2020, in a tissue culture room at Tufts University’s Allen Discovery Center, a developmental biologist named Douglas Blackiston watched a millimeter-wide clump of frog skin and heart cells push itself across a petri dish. Nothing about the clump had been programmed the way an engineer programs a robot. There was no chip, no motor, no code loaded onto silicon. The cells had been harvested from Xenopus laevis embryos, the African clawed frog long used in developmental biology, and shaped by hand under a microscope into a form that a computer had designed. The heart cells inside it contracted on their own rhythm, the way heart cells do in a living frog, and that contraction was enough to make the whole assembly crawl.
Blackiston worked with Michael Levin, who directs the Allen Discovery Center at Tufts, and Josh Bongard, a computer scientist at the University of Vermont who studies evolutionary robotics. The organism in the dish had no name yet that would stick in the press, but within weeks it had one: xenobot, after Xenopus. The paper describing it, published that January in the Proceedings of the National Academy of Sciences, carried an unglamorous title, “A scalable pipeline for designing reconfigurable organisms.” What it described was not unglamorous at all. It was a living machine designed entirely in simulation, then built out of cells that had never been genetically altered.
An Evolutionary Algorithm Designed the Body, Not a Human Engineer
The xenobot pipeline starts nowhere near a frog. It starts on the Vermont Advanced Computing Core, a supercomputing cluster at the University of Vermont nicknamed Deep Green. Bongard’s team ran an evolutionary algorithm on that cluster, testing millions of candidate body shapes made from simulated skin cells and simulated heart cells arranged in different configurations. Each candidate was scored on a simple task, usually locomotion or the ability to push an object toward a target. The algorithm bred the best-performing shapes against each other, mutated them, and ran the cycle again, generation after simulated generation, until it converged on forms no human designer would have sketched.
Those digital blueprints then went to Blackiston at Tufts, who used fine forceps and cauterizing tools to sculpt the actual shapes out of stem cells harvested from frog blastulae, the hollow ball of cells that forms a few days after fertilization. Skin cells were positioned in some places, cardiac cells in others, matching the computer’s design almost tissue for tissue. The result behaved the way the simulation predicted, walking, pushing, and in some configurations cooperating to move objects far larger than themselves. The genome inside every one of those cells was still 100 percent frog. Nothing had been edited. The only thing that changed was the anatomical context the cells found themselves arranged into.
The genome didn’t change. The anatomy did.
Cilia Turned Xenobots 2.0 Into Self-Healing Machines
The first xenobots relied on the twitch of cardiac muscle for movement, which made them slow and gave researchers limited control over direction. In March 2021, the same three-way collaboration between Tufts, the University of Vermont, and Harvard’s Wyss Institute published a follow-up in Science Robotics describing a version built differently. Instead of using embryonic cells sculpted into custom shapes, the team let adult frog skin cells self-assemble into spheroids in culture. Those spheroids came coated in cilia, the tiny hairlike structures that normally sweep mucus across the surface of a frog’s skin or a human airway. Repurposed for locomotion, the cilia beat in coordinated waves and drove the spheroids through liquid far faster than the muscle-powered version ever moved.
Xenobots 2.0 could also heal. Researchers sliced them nearly in half with a scalpel and watched the wound close on its own within minutes, the organism resuming normal movement afterward. That kind of self-repair is unremarkable in a frog. It is far more interesting in something built to spec by an algorithm, because it means the design pipeline had not stripped away the basic regenerative competence that skin cells carry with them regardless of what shape they’ve been arranged into. The cells still knew how to be an organism, even when the organism itself was new to biology.
Kinematic Replication: Xenobots That Build Copies of Themselves
The most consequential paper in the series appeared in PNAS in November 2021, again from Kriegman, Blackiston, Bongard, and Levin. It described xenobots that could reproduce, not by dividing the way a cell divides, and not by growing an embryo the way a frog does, but through a mechanism closer to what the mathematician John von Neumann described decades earlier in his theoretical work on self-replicating machines: kinematic replication, in which a structure uses raw material in its environment to physically assemble a copy of itself.
Left alone in a dish full of loose stem cells, spherical xenobots occasionally swept a few hundred of those free-floating cells into a pile using their cilia. Under the right conditions, that pile matured on its own into a new xenobot, structurally similar to its parent. Spheres, though, were inefficient at this. Sam Kriegman, then a PhD researcher under Bongard and now at Northwestern University, ran the design problem back through the Deep Green evolutionary algorithm, this time optimizing specifically for replication efficiency. Out of billions of simulated shapes, the algorithm converged on something that looked almost nothing like a sphere: a shape closer to Pac-Man, a body with a wide mouth-like opening that could scoop loose cells far more effectively than a rounded surface ever could.
The offspring wasn’t born. It was built by its parent, cell by loose cell.
Built physically to match that Pac-Man blueprint, the C-shaped xenobots swept up piles large enough to mature into fully formed second-generation xenobots, which then went on to build a third generation, and a fourth. The chain didn’t continue indefinitely. Each new generation tended to be smaller and less effective at gathering cells than the one before it, and the process petered out after a handful of cycles because the offspring eventually lacked the mass and cilia coverage needed to sweep up enough material. That ceiling is a real constraint on the platform, not a footnote. It means kinematic replication as demonstrated so far is a self-limiting process, not an open-ended one.
What the Numbering Actually Tracks
Popular coverage has attached version numbers to this work almost the way a phone manufacturer would, and the shorthand is useful as long as it’s read as journalistic packaging rather than an official designation from the Tufts or UVM labs. What the numbers really track is a progression in how much of the organism’s behavior originates from computed design versus emergent cell biology.
| Generation | Published | Cell source | Defining capability |
|---|---|---|---|
| Xenobots 1.0 | PNAS, Jan 2020 | Embryonic skin and cardiac cells | Muscle-driven locomotion from AI-designed shapes |
| Xenobots 2.0 | Science Robotics, Mar 2021 | Adult skin cell spheroids | Cilia-powered speed and self-healing |
| Kinematic replicators | PNAS, Nov 2021 | Embryonic stem cells, Pac-Man shape | Multi-generation self-replication |
| Anthrobots | Advanced Science, 2023 | Adult human tracheal cells | Proof the platform generalizes beyond frogs |
The Bioelectric Argument Underneath the Robotics Headlines
Levin’s interest in xenobots was never really about robotics. His lab has spent years studying how bioelectric signaling, the voltage gradients that form across cell membranes and the ion channels and gap junctions that let cells share those voltage states with their neighbors, directs the shape an organism builds during development and regeneration. Planaria flatworms in his lab have been made to grow heads with two brains or the wrong number of eyes by altering the voltage pattern across a wound site, with no genetic edit involved. Tadpole eyes, relocated to a frog’s tail, have been shown to wire into functional visual circuits using the bioelectric and neural signaling already present in that novel location.
Xenobots extend that argument into a much stranger context. Skin and heart cells carry a genome built to make a frog. Placed into an entirely new anatomical arrangement, with no frog body, no nervous system, and no developmental instructions telling them what they’re supposed to become, those same cells default to cooperative, coordinated, even self-repairing behavior. The genome didn’t specify “build a xenobot.” Something downstream of the genome, closer to the level of cell-to-cell electrical and mechanical signaling, filled in the rest.
Change the anatomical context, and the same genome builds a different machine.
That is the reason a bioelectricity publication cares about a frog-cell robot at all. The genetic code did not need editing to produce a novel living machine. The bioelectric and biomechanical signaling networks that already exist inside ordinary cells did the design work once the physical arrangement changed. Levin has argued in talks and papers that this points toward a layer of biological information processing sitting above the genome, a kind of software running on cellular hardware, and that learning to read and write it could eventually let researchers direct tissue repair, birth defect correction, or cancer reversion the way they now direct a robot, by specifying an outcome and letting cells find their own way there.
What Comes After Four Generations
The 2023 Anthrobot paper, led by Gizem Gumuskaya working with Levin, took the same design logic and applied it to adult human tracheal cells instead of frog cells, producing small multicellular structures that moved via cilia and, in some experiments, fused with damaged neural tissue in a dish and appeared to encourage regrowth across the gap. No frog was involved. No embryo was involved. The cells came from an adult airway, the kind harvested routinely in a hospital, which matters because it suggests the design pipeline built for xenobots is not a frog-specific curiosity. It is a method that appears to generalize to human tissue, raising the prospect of biological constructs built from a patient’s own cells for localized repair work, without the immune rejection risk that comes with foreign material.
The unresolved question is not whether xenobots can move or heal or make one more generation of themselves. They can, and the PNAS and Science Robotics papers document exactly how. The unresolved question is whether kinematic replication can be redesigned to avoid the shrinking-generation ceiling that stopped the Pac-Man replicators after a handful of cycles, and whether the Anthrobot line will ever be pushed toward a therapeutic application inside an actual patient rather than a dish. Levin’s lab has spent five years proving that cells will build novel bodies if the bioelectric and physical context calls for it. The next papers will show whether that context can be aimed at a specific human wound, rather than at a petri dish full of loose stem cells waiting to be swept up.