Xenobots

Xenobots are motile constructs, under a millimetre across, made by sculpting clumps of embryonic cells from the African clawed frog Xenopus laevis into shapes selected by an evolutionary algorithm running in a physics simulator. They carry no engineered hardware, no circuitry and no edited DNA. They move by beating cilia or by contracting embryonic heart muscle, and they run on yolk the cells carried out of the egg. The first were reported in 2020 by Sam Kriegman, Douglas Blackiston, Michael Levin and Josh Bongard, working between Tufts University and the University of Vermont.¹ The name joins the genus Xenopus to the back half of robot, and that second half accounts for most of the misunderstanding since.

§How they are made

The process has two stages that do not interact. In the first, an evolutionary algorithm searches a space of soft-body designs built from voxels of two kinds, passive tissue and contractile tissue. It scores each candidate in simulation on a task such as travelling in a straight line or pushing a particle, then filters the survivors for robustness, since a design that collapses under small perturbations will not survive transfer to tissue.

In the second stage a person builds the winning shape by hand. Animal cap tissue is dissected from blastula-stage frog embryos, dissociated, and left to reaggregate into a coherent ball. Microsurgical forceps and a cautery electrode then carve that ball to match the design, and in the original constructs a graft of cardiac progenitor tissue supplies the contractions that push the object along. Nothing is loaded into the finished construct afterwards: the design step delivers a body plan and stops. That is the first respect in which the word "robot" misleads.

A later generation dispensed with the surgery and the heart muscle. Animal cap cells left to themselves form spheroids that place their cilia on the outside and swim, using the same cilia that would otherwise sweep mucus across frog skin. These constructs repair themselves within minutes of being lacerated and mill through fields of loose particles. They can also be made to record an exposure to light by expressing a photoconvertible fluorescent protein, injected into the embryo as messenger RNA rather than written into the genome.² No genome editing is involved at any point; the cells carry the frog's ordinary genome in an unfamiliar setting.

Lifespan is set by the maternally loaded yolk platelets the cells inherited. Unfed, the ciliated constructs move for roughly ten days and then break down. In frog culture medium they have been kept alive for months, because the cells retain ordinary metabolic machinery, but none can find its own food.

§Development history

1. 1955
   Dissociated amphibian cells reaggregate
   Philip Townes and Johannes Holtfreter show that embryonic amphibian cells broken apart in culture reassemble and sort themselves by tissue type, establishing that architecture can arise from cells with no embryo around them.
2. 1966
   Kinematic self-reproduction formalized
   John von Neumann's posthumously edited work distinguishes a machine that copies a description of itself from one that merely assembles copies out of parts lying around it.
3. 2020
   First reconfigurable organisms
   Kriegman, Blackiston, Levin and Bongard report simulated designs built from Xenopus laevis skin and cardiac progenitor cells that move, push objects, and heal after being cut.
4. 2021
   Ciliated, self-assembling version
   Blackiston and colleagues report constructs that self-organize from skin cells alone and swim on their own cilia, removing the microsurgery step.
5. 2021
   Kinematic self-replication reported
   Parent constructs are shown to sweep loose stem cells into piles that mature into a further generation of swimmers, in a dish continuously supplied with fresh cells.
6. 2023
   Anthrobots
   Gumuskaya and colleagues in Levin's laboratory report motile spheroids that self-assemble from adult human tracheal epithelial cells.

§Kinematic self-replication

The 2021 report that xenobots "self-replicate" is the most misread result in the field.³ The observed process is mechanical. A parent construct swimming through a dish strewn with dissociated stem cells pushes them into heaps; a heap that reaches a threshold size compacts, grows cilia, and swims off as a new construct. An evolutionary search found a C-shaped parent, widely described in coverage as AI-designed, that gathers cells more efficiently than a sphere and sustains the cycle for more rounds.

No genome is read or written and no template is used. The offspring are spheroids that do not inherit the parent's shape, so the feature that made the parent a good replicator is lost in the first generation. An experimenter has to keep restocking the dish with fresh cells, and without that supply the process stops immediately. This is replication in von Neumann's kinematic sense, the assembly of a copy from parts already present, rather than in the biological sense of heredity with variation.⁴

Not replication in the biological sense

Constructs that genuinely propagate a design through a population, such as a Gene drives, do it by copying DNA, which is the mechanism missing here. Without heredity there is no variation for selection to act on, so the process cannot evolve. The runaway scenarios of the Grey goo literature also require a closure over materials and energy that nothing in this work approaches.

§Anthrobots and human cells

The same laboratory has reported an equivalent construct made from adult human cells. Anthrobots self-assemble from donated tracheal epithelial cells, which under the right culture conditions turn their cilia outward and become motile spheroids ranging from tens to a few hundred micrometres across.⁵ Unlike constructs built on Induced pluripotent stem cells, they are made without reprogramming, and unlike Organ bioprinting or Decellularized scaffolds, without an external scaffold. The only thing imposed on the cells is the medium they sit in.

In the reported experiment, clusters of anthrobots were placed across a scratch in a monolayer of cultured human neurons, and regrowth followed across the gap in a way that untreated controls did not show. That is a result in a dish. No mechanism has been established and no animal has been treated. The step from closing a scratch in culture to repairing tissue in a person is one that most of regenerative medicine has never managed.

§The morphogenesis claim

The scientifically interesting argument attached to this work concerns development rather than robotics. Levin holds that the genome specifies cellular hardware rather than anatomy directly, that cells and tissues coordinate through bioelectric signalling toward target morphologies, and that a group of cells removed from the signalling context of an embryo will settle into a different but stable body plan.⁶ On that reading a xenobot is a frog cell collective exercising an option its genome always contained, rather than a machine made of tissue. Levin and Daniel Dennett have pressed a stronger version of the framing, in which goal-directedness is a property found at every scale of biology rather than only in nervous systems.⁷

Supporting work outside the xenobot programme comes from planarian flatworms. After a brief disruption of gap-junction signalling a minority regenerate two heads, and those animals go on yielding two-headed worms through later amputations in plain water, with no change to their DNA.⁸ The result bears on Limb regeneration and on what a genome does and does not determine.

How much does the construct actually show

Critics accept the observations and dispute the interpretation. Dissociated amphibian cells were known to reaggregate and sort by tissue type in the 1950s, and ciliated epithelium swims because that is what ciliated epithelium does.⁹ On the sceptical reading, a xenobot is a motile explant with a haircut, and the language of body plans, goals and reconfigurable organisms is doing work the data do not require. Levin's side replies that a stable, reproducible, functional morphology absent from the frog's life cycle is exactly what needs explaining. The two camps disagree about the explanatory frame; neither disputes the images.

§Limitations

Coverage repeats two misreadings almost without exception. The first concerns scale. These are sub-millimetre objects, orders of magnitude larger than the constructions in the Medical nanorobots tradition and comparable to the larger devices in Medical microrobots. The second is that they are machines. They contain no sensor, actuator, controller or program. The evolutionary search runs on a computer before any cells are touched, and the resulting object cannot be reprogrammed, updated, or told what to do.

The transfer from simulation to tissue is loose. The simulator treats cells as voxels of passive or contractile material and omits adhesion, signalling, and the tendency of tissue to remodel itself, so built constructs approximate their designs rather than instantiate them. Batch variation is substantial, as it is with Organoids, and nothing can be steered: a shape biases where a swimmer tends to go, so motion is directed only in a statistical sense.

No therapeutic application exists. The proposals in circulation, such as clearing microplastics from water, scraping plaque from vessels, or carrying a payload in the manner of Targeted drug delivery, are speculative in the strict sense that none has been attempted in an animal. A construct that runs out of fuel in days, cannot be located once released, and consists of foreign frog cells has no obvious route into a body.

§Risk and governance

The mainstream assessment is that the biosafety risk is low, and it rests on specific properties rather than general reassurance. Xenobots cannot feed themselves, cannot reproduce without cells supplied to them, and are built from vertebrate cells that die outside a controlled medium. That places them below the containment concerns behind the safeguards in Genetic code expansion and recoding, and far below those about Mirror life. Invoking the Precautionary principle here amounts to an argument about a research programme whose future forms cannot be specified; it says nothing about the constructs already made.

The harder questions are definitional. Whether these things are organisms, and what would be owed to them if they were, runs into the same lack of an agreed test that troubles Machine consciousness, though with less urgency than for cerebral organoids, since a xenobot contains no neurons at all. Regulatory categories fit badly. A construct of unmodified animal cells is not a genetically modified organism, a medical device or a laboratory animal, the same gap that Stem-cell-based embryo models have opened from the other direction. The technique is also cheap and needs no gene synthesis, so it sits outside most of the machinery built for Dual-use research of concern.

§Outlook

In the near term the work is worth more as an experimental system than as a product. It offers a way to ask what a group of cells builds when the embryo's instructions are removed, on a timescale of days and at a cost that permits many replicates. That question sits upstream of Tissue engineering and of every attempt at Lab-grown organs. Whether the constructs themselves ever do useful work is a separate and longer bet, which would need a food supply, a means of control, and a reason to prefer a living object to an engineered one.

The claim that would matter most if it held is also the hardest to test. If anatomy is a reconfigurable target rather than a fixed genomic output, then instructing cells at the level of pattern, instead of editing the genes beneath it, becomes a legitimate strategy for regeneration. Frog cells arranged into something the frog never makes show that the option exists. They do not yet show that anyone knows how to specify which option is taken.

§See also

• Organoids
• Stem-cell-based embryo models
• Limb regeneration
• Tissue engineering
• Medical microrobots
• Medical nanorobots
• Machine consciousness
• Mirror life

§References