Xenobots: revision 1 to 2 — Future Human Wiki

Revision 1

Revision 2

categories: ["bodies"]

tags: ["synthetic biology", "morphogenesis", "biorobotics", "regeneration", "self-assembly", "michael levin"]

summary: "Sub-millimetre motile constructs assembled from frog embryonic cells and shaped by an evolutionary algorithm, classed as neither robots nor organisms in any settled sense."

summary: "Sub-millimetre motile constructs assembled from frog embryonic cells and shaped by an evolutionary algorithm, classed as neither robots nor organisms in any settled sense."

updated: "2026-07-27"

updated: "2026-10-29"

humanEvidence: "No xenobot has been placed in an animal or a person; the human-cell version, anthrobots, has been tested only in culture dishes, including on sheets of cultured human neurons."

humanEvidence: "No xenobot has been placed in an animal or a person; the human-cell version, anthrobots, has been tested only in culture dishes, including on sheets of cultured human neurons."

access: "Not a product. Built in a small number of academic laboratories from frog embryos or donor-derived human airway cells; none are sold, licensed, or approved for any use."

access: "Not a product. Built in a small number of academic laboratories from frog embryos or donor-derived human airway cells; none are sold, licensed, or approved for any use."

reversibility: "reversible"

}

```

**Xenobots** are motile constructs, under a millimetre across, made by sculpting clumps of

**Xenobots** are motile constructs, under a millimetre across, made by sculpting clumps of embryonic

embryonic cells from the African clawed frog *Xenopus laevis* into shapes selected by an

cells from the African clawed frog *Xenopus laevis* into shapes selected by an evolutionary

evolutionary algorithm running in a physics simulator. They carry no engineered hardware, no

algorithm running in a physics simulator. They carry no engineered hardware, no circuitry and no

circuitry, and no edited DNA; motion comes from beating cilia or from contracting embryonic heart

edited DNA. They move by beating cilia or by contracting embryonic heart muscle, and they run on

muscle, and the energy comes from yolk the cells carried out of the egg. The first were reported in

yolk the cells carried out of the egg. The first were reported in 2020 by Sam Kriegman, Douglas

2020 by Sam Kriegman, Douglas Blackiston, Michael Levin and Josh Bongard, working between Tufts

Blackiston, Michael Levin and Josh Bongard, working between Tufts University and the University of

University and the University of Vermont.[^kriegman2020] The name joins the genus *Xenopus* to the

Vermont.[^kriegman2020] The name joins the genus *Xenopus* to the back half of *robot*, and that

back half of *robot*, and that second half has caused most of the trouble since.

second half accounts for most of the misunderstanding since.

## How they are made

The pipeline has two halves that never touch. In the first, an evolutionary algorithm searches a

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 —

space of soft-body designs built from voxels of two kinds, passive tissue and contractile tissue. It

scoring each candidate in simulation on a task such as travelling in a straight line or pushing a

scores each candidate in simulation on a task such as travelling in a straight line or pushing a

particle, then filtering the survivors for robustness, since a design that collapses under small

particle, then filters the survivors for robustness, since a design that collapses under small

perturbations will not survive transfer to tissue.

In the second half a person builds the winning shape by hand. Animal cap tissue is dissected from

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

object along. Nothing is loaded into the finished construct afterwards: the design step delivers a

body plan and stops, which is the first respect in which the word "robot" misleads.

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 — the same cilia that would

themselves form spheroids that place their cilia on the outside and swim, using the same cilia that

otherwise sweep mucus across frog skin — and swim. These constructs repair themselves within

would otherwise sweep mucus across frog skin. These constructs repair themselves within minutes of

minutes of being lacerated, mill through fields of loose particles, and can be made to record an

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.[^blackiston2021] Nothing here involves

messenger RNA rather than written into the genome.[^blackiston2021] No

[[crispr-cas9|genome editing]]: the cells carry the frog's ordinary genome in an unfamiliar setting.

[[crispr-cas9|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

constructs move for roughly ten days and then break down. In frog culture medium they have been kept

alive for months, since the cells retain ordinary metabolic machinery. None can find its own food.

alive for months, because the cells retain ordinary metabolic machinery, but none can find its own

food.

## Development history

## Kinematic self-replication

The 2021 report that xenobots "self-replicate" is the most misread result in the

field.[^kriegman2021] What was observed is mechanical. A parent construct swimming

field.[^kriegman2021] The observed process is mechanical. A parent construct swimming through a dish

through a dish strewn with dissociated stem cells pushes them into heaps; a heap that reaches a

strewn with dissociated stem cells pushes them into heaps; a heap that reaches a threshold size

threshold size compacts, grows cilia, and swims off as a new construct. An evolutionary search found

compacts, grows cilia, and swims off as a new construct. An evolutionary search found a C-shaped

a C-shaped parent, widely described in coverage as AI-designed, that gathers cells more efficiently

parent, widely described in coverage as AI-designed, that gathers cells more efficiently than a

than a sphere and sustains the cycle for more rounds.

sphere and sustains the cycle for more rounds.

Nothing is copied. No genome is read or written, no template is used, and the offspring are

No genome is read or written and no template is used. The offspring are spheroids that do not

spheroids that do not inherit the parent's shape, so the very feature that made a good replicator is

inherit the parent's shape, so the feature that made the parent a good replicator is lost in the

lost in the first generation. The dish must be restocked with fresh cells by an experimenter; take

first generation. An experimenter has to keep restocking the dish with fresh cells, and without that

that away and the process stops immediately. This is replication in von Neumann's kinematic sense —

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

assembly of a copy from parts already present, rather than in the biological sense of heredity with

variation.[^vonneumann1966]

> [!caution] Not replication in the biological sense

> Constructs that genuinely propagate a design through a population, such as a [[gene-drive]], do it

> by copying DNA — the mechanism absent here. Without heredity there is no variation for selection to

> by copying DNA, which is the mechanism missing here. Without heredity there is no variation for

> act on, so the process cannot evolve, and the runaway scenarios of the [[grey-goo]] literature

> selection to act on, so the process cannot evolve. The runaway scenarios of the [[grey-goo]]

> require a closure over materials and energy that nothing in this work approaches.

> 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.[^gumuskaya2023] They are made without reprogramming, unlike constructs built on

across.[^gumuskaya2023] Unlike constructs built on [[induced-pluripotent-stem-cells]], they are made

[[induced-pluripotent-stem-cells]], and without an external scaffold, unlike [[organ-bioprinting]]

without reprogramming, and unlike [[organ-bioprinting]] or [[decellularized-scaffolds]], without an

or [[decellularized-scaffolds]]. Nothing is imposed on the cells except the medium they sit in.

external scaffold. The only thing imposed on the cells is the medium they sit in.

In the reported experiment, clusters of anthrobots placed across a scratch in a monolayer of

In the reported experiment, clusters of anthrobots were placed across a scratch in a monolayer of

cultured human neurons were followed by regrowth across the gap that untreated controls did not

cultured human neurons, and regrowth followed across the gap in a way that untreated controls did

show. That is a result in a dish. No mechanism has been established and no animal has been treated,

not show. That is a result in a dish. No mechanism has been established and no animal has been

and the distance between closing a scratch in culture and repairing tissue in a person is the one

treated. The step from closing a scratch in culture to repairing tissue in a person is one that most

most of regenerative medicine has never crossed.

of regenerative medicine has never managed.

## The morphogenesis claim

The scientifically interesting argument attached to this work is not about robotics at all. Levin

The scientifically interesting argument attached to this work concerns development rather than

holds that the genome specifies cellular hardware rather than anatomy directly, that cells and

robotics. Levin holds that the genome specifies cellular hardware rather than anatomy directly, that

tissues coordinate through bioelectric signalling toward target morphologies, and that a group of

cells and tissues coordinate through bioelectric signalling toward target morphologies, and that a

cells removed from the signalling context of an embryo will settle into a different but stable body

group of cells removed from the signalling context of an embryo will settle into a different but

plan.[^levin2021] On that reading a xenobot is not a machine made of tissue but a frog cell

stable body plan.[^levin2021] On that reading a xenobot is a frog cell collective exercising an

collective exercising an option its genome always contained. Levin and Daniel Dennett have pressed a

option its genome always contained, rather than a machine made of tissue. Levin and Daniel Dennett

stronger version of the framing, in which goal-directedness is a property found at every scale of

have pressed a stronger version of the framing, in which goal-directedness is a property found at

biology rather than only in nervous systems.[^dennett2020] Supporting work outside the xenobot

every scale of biology rather than only in nervous systems.[^dennett2020]

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.[^durant2017] The result bears on

[[limb-regeneration]] and on what a genome does and does not determine.

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.[^durant2017] The result bears on [[limb-regeneration]] and on what a genome does and does not

determine.

> [!debate] 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.[^townes1955] On the sceptical reading, a xenobot is a motile

> that is what ciliated epithelium does.[^townes1955] On the sceptical reading, a xenobot is a

> explant with a haircut, and the language of body plans, goals and reconfigurable organisms is

> motile explant with a haircut, and the language of body plans, goals and reconfigurable organisms

> doing work the data do not require. The reply from Levin's side is that a stable, reproducible,

> 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 precisely what needs explaining. The

> functional morphology absent from the frog's life cycle is exactly what needs explaining. The two

> disagreement is about the explanatory frame, not the images.

> camps disagree about the explanatory frame; neither disputes the images.

## Limitations

Two misreadings are near-universal in coverage. These are sub-millimetre objects, orders of

Coverage repeats two misreadings almost without exception. The first concerns scale. These are

magnitude larger than the constructions in the [[medical-nanorobots]] tradition and comparable to

sub-millimetre objects, orders of magnitude larger than the constructions in the

the larger devices in [[microrobots-in-medicine]]; and they are not machines, containing no sensor,

[[medical-nanorobots]] tradition and comparable to the larger devices in

no actuator, no controller, and no program. The evolutionary search runs on a computer before any

[[microrobots-in-medicine]]. The second is that they are machines. They contain no sensor, actuator,

cells are touched, and the resulting object cannot be reprogrammed, updated, or told what to do.

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, omitting adhesion, signalling, and the tendency of tissue to remodel itself,

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, in the manner familiar from [[organoids]], and nothing can be steered: motion is

substantial, as it is with [[organoids]], and nothing can be steered: a shape biases where a swimmer

directed only in the statistical sense that a shape biases where a swimmer tends to go.

tends to go, so motion is directed only in a statistical sense.

No therapeutic application exists. Proposals in circulation — clearing microplastics from water,

No therapeutic application exists. The proposals in circulation, such as clearing microplastics from

scraping plaque from vessels, carrying a payload in the manner of [[targeted-drug-delivery]] — are

water, scraping plaque from vessels, or carrying a payload in the manner of

speculative in the strict sense that none has been attempted in an animal. A construct that runs out

[[targeted-drug-delivery]], are speculative in the strict sense that none has been attempted in an

of fuel in days, cannot be located once released, and consists of foreign frog cells has no obvious

animal. A construct that runs out of fuel in days, cannot be located once released, and consists of

route into a body.

foreign frog cells has no obvious route into a body.

## Risk and governance

The mainstream assessment of biosafety risk is that it is low, for a specific reason rather than a

The mainstream assessment is that the biosafety risk is low, and it rests on specific properties

reassuring one: xenobots cannot feed themselves, cannot reproduce without cells supplied to them,

rather than general reassurance. Xenobots cannot feed themselves, cannot reproduce without cells

and are built from vertebrate cells that die outside a controlled medium. That places them below the

supplied to them, and are built from vertebrate cells that die outside a controlled medium. That

containment concerns motivating the safeguards in [[recoded-organisms]], and far below those about

places them below the containment concerns behind the safeguards in [[recoded-organisms]], and far

[[mirror-life]]. Invoking the [[precautionary-principle]] here is an argument about a research

below those about [[mirror-life]]. Invoking the [[precautionary-principle]] here amounts to an

programme whose future forms cannot be specified, not about the constructs already made.

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, meets the same absence of an agreed test that troubles

them if they were, runs into the same lack of an agreed test that troubles

[[machine-consciousness]] — with less urgency than for cerebral organoids, since a xenobot contains

[[machine-consciousness]], though with less urgency than for cerebral organoids, since a xenobot

no neurons at all. Regulatory categories fit badly: a construct of unmodified animal cells is not a

contains no neurons at all. Regulatory categories fit badly. A construct of unmodified animal cells

genetically modified organism, not a medical device, and not a laboratory animal, the same gap

is not a genetically modified organism, a medical device or a laboratory animal, the same gap that

[[synthetic-embryos]] 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]].

## Outlook

The near-term value of the work is as an experimental system rather than a product. It offers a way

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, a question upstream of [[tissue-engineering]] and

days and at a cost that permits many replicates. That question sits upstream of

of every attempt at [[lab-grown-organs]]. Whether the constructs themselves ever do useful work is a

[[tissue-engineering]] and of every attempt at [[lab-grown-organs]]. Whether the constructs

separate and longer bet: the missing pieces are a food supply, control, and a reason to prefer a

themselves ever do useful work is a separate and longer bet, which would need a food supply, a means

living object to an engineered one.

of control, and a reason to prefer a living object to an engineered one.

The claim that would matter most if it held is the one hardest to test. If anatomy is a

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 demonstrate that the option exists. They are

Frog cells arranged into something the frog never makes show that the option exists. They do not yet

not yet evidence that anyone knows how to specify which option is taken.

show that anyone knows how to specify which option is taken.

## See also