Listen: https://serve.buildtall.systems/abf2e85a9e9c6213f893eda65ff5dc04fa630d63e3075daf36dc57d31173044a.mp3
Source: https://serve.buildtall.systems/619559f78659974148e2d7d9e0c1588e598d2ac05f2ce845213367088e2e846f.pdf
This episode covers Microcosmos: Reimagining Artificial Life for the GPU Era, a paper by Mark Tensen and colleagues at the Artificial Life Institute in Japan, with Kenneth O. Stanley of Lila Sciences among the co-authors.
The paper asks what keeps artificial life from reaching the open-ended evolutionary complexity seen in nature, and proposes Microcosmos, a simulation engine in which artificial creatures are elastic filament chains that swim, fold and forage in a two-dimensional viscous fluid. The engine is built for modern graphics processors and is end-to-end differentiable, which means the effect of each design parameter on the outcome can be calculated and used to guide a search.
The front matter gives the work as a full paper submission and as an arXiv preprint, identified as version 1 and dated 3 July 2026. According to OpenAlex, it is a conference paper published on 1 August 2026, and the index lists no retraction. The paper itself names no conference, so the two sources differ on the venue: the document presents itself as an arXiv preprint, while the index gives a conference date.
According to OpenAlex, the paper has been cited 0 times as of 19 October 2026. The index gives no trend by year, and the work is very recent.
The evidence is four simulation experiments. The locomotion check is qualitative: it compares the simulated creatures with a known physical prediction rather than with measured organisms. The folding experiment covers all ten digit classes, but the authors themselves call the reproductions imperfect. The scaling result comes from one graphics processor and runs up to half a million particles. The authors also state limits of their own: their handling of collisions between filaments is approximate, and outside a certain range of fluid thickness their coupling may produce artifacts. The extracted text includes affiliations but no funding statement. The authors point to released code and supplementary videos.
The paper opens with a question: "What is preventing artificial life (ALife) from achieving the kind of open-ended evolutionary complexity we observe in nature?" The authors note that many of the open problems set out in 2000 by Bedau and colleagues remain unsolved. They argue that the field is held back by "both a lack of physical grounding in our simulation substrates and insufficient computational scalability to support evolutionary search at scale."
The paper sorts existing tools into two groups. Abstract, rule-based substrates are efficient to evolve, but the authors say they are disconnected from the physical principles studied in fields such as biophysics and fluid dynamics. More physically grounded simulators have the opposite problem. In the authors' account, they are computationally intractable for large evolutionary searches, operate at too high a level of abstraction, or are designed for reinforcement learning rather than for the open-ended dynamics of life. Progress, they write, demands simulators that are "physically grounded enough to be credible, efficient enough to support evolutionary search at scale, and behaviorally rich enough to inspire."
The authors look to the natural world for two sources of guidance. The first is locomotion at the microscale, where organisms move through viscous fluid and inertia is negligible. The diversity of swimming gaits, body plans and collective behaviors in microorganisms comes from the interplay between elastic body mechanics and the surrounding fluid. The second is protein folding, in which linear chains of amino acids collapse into three-dimensional structures that determine biological activity.
In both cases, the paper says, the building blocks are filaments: chains of connected units with simple nearest-neighbor topology. The authors call the shared structure "not a coincidence." Filamentous organization, which runs from bacterial flagella to cytoskeletal networks to nematode body plans, is among the oldest and most widespread forms of biological complexity. Nearest-neighbor connectivity also makes filament simulation highly scalable.
On this basis the paper presents Microcosmos, a simulation engine that takes filaments as its core abstraction for artificial life. In the model, artificial lifeforms are elastic filament chains living in a two-dimensional viscous fluid. Their genetic encodings specify both their form and their behavior. Within this world, the creatures swim, fold into precise shapes and forage across chemical gradients.
The authors report a range of experiments to support these claims. They say the fluid simulation produces jellyfish, cilia and tadpole gaits that are consistent with known physical constraints. They show that the whole system is differentiable by training filaments to fold into target shapes using gradient descent. Using neuroevolution and quality-diversity search, they find a wide range of swimming strategies discovered automatically. Finally, they report that runtime increases linearly with the number of particles.
The paper makes a priority claim:
To our knowledge, this is the first system to combine differentiable flexible filament simulation with a resolved fluid solver for both morphology and controller optimization.
The authors release Microcosmos as an open platform, "in the hope that it serves as a foundation for the wider artificial life community to build upon."
The paper's related work section begins with the field of artificial life, the study of lifelike behavior produced inside computers. The authors say that artificial life research has explored "a diverse range of computational substrates in pursuit of open-ended complexification." In plain terms, a substrate is the underlying medium the simulated life runs on, and open-ended complexification means systems that keep producing new and more intricate forms instead of leveling off.
The paper sorts earlier work into two families. The first is abstract computational worlds. It names cellular automata, which are grids of cells that change state according to simple local rules, along with Lenia, a continuous relative of that idea, Tierra and Avida, which are digital evolution systems where programs compete and reproduce, and two further systems, BFF and Chromaria. The authors write that these systems "show that rich evolutionary dynamics can emerge from minimal rules."
The second family is particle-based systems. These include Boids, a classic model of flocking birds, Particle Life, and Particle Lenia. Here the paper says that these systems "demonstrate that self-organizing collective behaviors can likewise emerge from simple local interactions," meaning that group patterns form from each unit responding only to its neighbors.
The authors then note a shared weakness. "All of these substrates remain disconnected from physical reality," they write, which limits the embodied behaviors they can support. Embodied means the creatures have bodies that must move and act within a physical environment.
The paper next turns to physically grounded, embodied simulators. The authors say these show that worlds built on collisions between bodies can produce diverse behaviors. The catch is cost: "such approaches are often computationally intractable," meaning they demand more computing than is practical to run.
Other systems take a different route to speed. Some use agent-based frameworks, in which many individual units follow rules, and run them on graphics processing units, the chips built to carry out large numbers of calculations at once. Others are designed to scale indefinitely. According to the paper, these systems stay tractable, meaning practical to run, but they "operate at higher levels of abstraction," so they simplify away much of the physical detail.
The passage closes on the authors' claimed answer to this gap.
Microcosmos addresses these limitations by grounding artificial life in viscous fluid dynamics, producing embodied creatures whose behaviors are physically plausible, visually compelling and computationally tractable at scale.
Viscous fluid dynamics refers to the physics of thick, sticky liquids, the setting the paper chooses for its creatures.
The bio-inspiration behind the work begins with the microscopic world, which the authors describe as a compelling substrate for artificial life research. Filamentous structures, including chains of cells, polymers, flagella, and cytoskeletal networks, are among the most ancient forms of biological organization. The authors argue that elongated elastic bodies immersed in viscous fluids offer "a minimal yet expressive design space in which evolution has produced remarkable functional diversity," with examples running from bacterial flagella to nematode body plans.
The authors then set out the physical constraints of this regime. At low Reynolds numbers, where thick viscous forces dominate over inertia, the Navier-Stokes equations that govern fluid flow "become time-reversible." The consequence, drawn from Purcell's scallop theorem, is that "reciprocal motion cannot produce net locomotion in Stokes flow." In other words, a body that simply repeats the same back-and-forth shape changes will get nowhere. To move, a body has to deform in a way that cannot be undone by running the sequence backward.
The paper points to several biological solutions to this problem. One is the undulatory wave of the nematode C. elegans. Another is the propagating transverse wave that Taylor analyzed. A third is the asymmetric ciliary stroke that Purcell also described. Each one breaks the symmetry that would otherwise trap a swimmer in place.
From these constraints, the authors conclude that low-Reynolds-number locomotion is "a natural testbed for studying how physics shapes the evolution of morphology and behavior." The passage closes on the authors' proposed approach, which they describe as
a fully coupled lattice Boltzmann fluid simulation, designed from the ground up for evolutionary search over body plans and behaviors.
Computer graphics, the field concerned with building believable virtual worlds, has long shaped artificial life. The paper describes that core aim as "the efficient synthesis of believable virtual environments", and it argues that breakthroughs in graphics have repeatedly driven breakthroughs in the study of artificial life.
It gives three early examples. Boids was first a graphics technique for animating flocks, and it became a foundational model for emergent collective behavior, in which simple local rules produce group patterns that no single agent plans. Sims' virtual creatures applied rigid-body physics, the mechanics of solid objects that move and collide, from computer animation to a simulated process of biological evolution.
The paper says this trend continues. It points to differentiable graphics tools, DiffTaichi and DiffPD, which allow gradient-based optimization of soft bodies and fluid-structure interactions. Gradient-based optimization adjusts a design step by step by following the slope of an error measure toward a better result. The paper also notes the use of voxel-based physics, which models a body as a grid of small cubes, to design soft robots, citing work by Cheney and colleagues and by Kriegman and colleagues. More recently, advances in real-time physics simulation have made possible reinforcement learning benchmark frameworks, HydroGym and FluidGym, in which a learning agent improves its control of simulated fluid flows through trial and error.
Microcosmos, the paper says, continues this lineage. It draws on fluid and soft-body simulations from graphics that run efficiently on GPUs, the parallel processors built for graphics work, in order to "ground artificial life in the mechanics that govern life at the microscale."
The paper then sets out the engine's design. It has two coupled components. Filaments, which represent the artificial lifeforms, are built as graphs, where each node's state records its position in continuous space. Fields represent the environment as discrete grids that store information about the world, such as the state of the fluid. The two are overlaid and interact as a dynamical system. The paper presents this as a middle ground between grid-based approaches, such as cellular automata, and particle-based approaches, such as Particle Life.
The whole simulator is implemented in JAX, a library that enables GPU acceleration and end-to-end automatic differentiation, meaning that gradients can be computed through the whole simulation. The paper states that the filament-field design is motivated largely by scalability. The engine must support large populations of interacting filaments over evolutionary timescales, so, in its words, "every algorithmic choice avoids quadratic scaling with the number of nodes." Quadratic scaling means that doubling the number of nodes roughly quadruples the work. The fluid solver, for example, needs only local updates. The authors say these choices make the simulation scalable and fully parallelizable.
The third strand of prior work joins design optimization to fluid dynamics. Filaments have been simulated in resolved fluids without any optimization, as in the work of Tekinalp and colleagues and of Tian and colleagues. The paper's objection is that these systems are not differentiable, so "any search over morphology must rely on gradient-free methods." A differentiable system can report how its output changes when each design parameter is nudged, which lets an optimizer adjust many parameters at once instead of trying variations blindly.
Other work has made filament mechanics differentiable. Bergou and colleagues, and Stuyck and Chen, enabled gradient-based optimization of rest shapes and stiffness. The paper points out the missing piece: "without hydrodynamic coupling the optimized morphologies have no notion of locomotion or drag." Hydrodynamic coupling means the filament pushes on the surrounding fluid while the fluid pushes back, so the shape's effect on movement is part of the calculation.
A further group has paired differentiable fluid solvers with design optimization, including work by Ma and colleagues, Lee and colleagues, and Fan and colleagues. According to the paper, these systems "use rigid or beamlike bodies and target engineering inverse problems rather than morphological evolution."
The paper then compares the three closest systems, each of which misses one of the three ingredients. SophT, by Tekinalp and colleagues, couples a resolved fluid to Cosserat rods, which are slender elastic rods that model bending and twisting, but it is not differentiable. PyElastica, by Zhang and colleagues, provides Cosserat rods with optimization through reinforcement learning, a trial-and-error training method, but it approximates drag locally rather than solving the flow around the body. Diff-FlowFSI, by Fan and colleagues, is fully differentiable with resolved flow, but its beam elements lack "the topological flexibility needed for morphological evolution."
The paper places its own work in the gap these systems leave open. It aims at gradient-based optimization of flexible filament morphology, with fluid coupling that resolves the flow, so that the shape of a filament can change while the simulation keeps track of how it moves through the fluid.
The methods turn to filaments, the thread-like structures the simulator is built from. The paper notes that filamentous structures span the micro and nanoscale, from bacterial flagella and cytoskeletal networks that drive locomotion at low Reynolds numbers, to proteins that fold one-dimensional sequences into three-dimensional volumes, to contractile muscle filaments. The authors model each filament as "discrete elastic rods that resist both stretching and bending," and say these rods can be extended to have loops and branches through a graph representation.
Filaments do not collide with one another directly. Instead, intersections are avoided through a grid-based steric repulsion field, rather than handling collisions particle by particle. The authors are candid about the trade-off:
The steric field does not guarantee exact collision dynamics; however, the benefits of scaling are worth the loss in accuracy.
Two footnotes place this design in context. The first says that in fluid-solid interaction research this is a common paradigm, in which grid-based fields are called the Eulerian component and continuous-valued graphs, which follow the material as it moves, are called the Lagrangian component. The second says the same idea already exists in artificial life. In models of slime molds, agents move through continuous space but deposit pheromones onto a discrete grid, where the chemicals diffuse and other agents read them to guide their behavior.
Each segment of a filament carries its own orientation angle, which the paper calls the material frame of the edge. This follows the Cosserat rod formulation, a standard way of describing rods that track both position and orientation. Physics settings are assigned edge by edge rather than for the whole body. These include a rest length, a rest bending angle, and optional stiffness coefficients that control how strongly bending, shear and stretching are enforced. Together they define a creature's body shape, and they are the quantities that both gradient-based and evolutionary optimization adjust.
The dynamics are computed with Position-Based Dynamics, an iterative method that repeatedly nudges node positions and edge orientations until geometric constraints are met. The authors give two reasons for choosing it. It is "unconditionally stable regardless of timestep," and its operations are simple arithmetic that the JAX library can differentiate through directly. Implicit solvers, by contrast, would require differentiating through a linear solve, which is harder to do.
Adapting Cosserat constraints to this framework, for the two-dimensional case, takes each iteration through two phases. The first is a bending pass, which moves the orientation angles toward the rest bending angle. For two adjacent edges, the bending error is the difference between their two angles minus the rest angle. The correction is applied to both edges symmetrically, each receiving half the stiffness-weighted error in opposite directions. The second is a position pass, which enforces connectivity by correcting stretch and shear errors using the updated orientations.
In the position pass, an edge joins two nodes, and its vector is the difference between their positions. A material director is derived from the edge's orientation angle. The stretch constraint pulls the edge length toward its rest length along the geometric direction of the edge. To separate stretch from shear, the total positional correction blends this stretch constraint with the full Kirchhoff-Love error, which measures how far the edge vector is from the rest length along the material director. A shear stiffness parameter, between zero and one, sets the balance between the two. The resulting correction is applied to the two nodes symmetrically, one moved forward and the other backward.
The paper next defines the fields that make up its simulated world. The primary one is fluid velocity, a vector field that persists through the simulation, so every point in the space carries a direction and speed of flow. The authors compute it with the Lattice Boltzmann Method, which tracks packets of particle-like values on a fixed grid rather than solving the fluid equations directly. They use the D2Q9 variant, which works in two dimensions and moves values among nine possible directions at each grid point. The authors chose the method for practical reasons. In their words, it is "naturally parallel on GPU", because "each lattice site updates independently during the streaming step", so a graphics processor can work on many sites at once. They add that it is compatible with spatial domain decomposition, meaning the simulated space can be split into regions that are handled separately. The paper places the method at the mesoscopic scale, between molecular dynamics, which follows individual molecules, and direct Navier-Stokes solvers, which compute fluid motion from the governing equations. The method also handles moving filament boundaries without remeshing, so the grid does not have to be rebuilt as the swimming filament moves. Viscosity, the thickness or resistance of the fluid, is set by one relaxation parameter, tau, which the paper says makes it "straightforward to sweep across physical regimes", that is, to run the same simulation across many fluid conditions with little effort. Filament nodes must also avoid passing through one another. Rather than checking every pair of nodes, which would scale with the square of their number, the authors deposit node positions onto a grid, smooth them with a Gaussian filter applied through fast Fourier transform convolution to form a density field, and use the gradient of that field, the direction in which density rises most steeply, to generate repulsive forces that push nodes apart. For the neuroevolution experiments, the authors add a second, persistent scalar energy field made of Gaussian packets, which are bell-shaped bumps of energy placed at random positions. This field gives the evolving creatures a chemotaxis target, a gradient toward which they can learn to move.
Fluid and filaments interact through the Immersed Boundary Method, or IBM, a technique the authors credit to earlier work by Peskin. The exchange runs in both directions. First, the filaments' nodes sample the local fluid velocity to determine their own drag. Second, the agent, meaning the simulated creature, "broadcasts its own velocity back onto the grid."
The method does not treat the filament as a sharp edge. The interactions are smoothed over a small neighborhood, which the authors call a diffuse interface. In their words, the method "converts single node positions into a continuous, approximated solid body."
That smoothing brings a known problem. A diffuse boundary is not fully solid, so fluid can slip through it. The authors name this as "the challenge of fluid leaking through the filaments," and say it is especially the case with a zero-thickness filament, one that has no width of its own. To deal with it, they draw on several established practices from the lattice Boltzmann and immersed boundary literature. The first is Multi-Direct Forcing, credited to Wang and colleagues. The second is Two-Relaxation-Time, credited to Ginzburg and colleagues. The third is a structural workaround. Rather than leave the filament as a single line, the authors broadcast it into a two-layer version of itself, which they describe as a way to "artificially amplify a zero-thickness filament into a solid object." They credit this idea to Mittal and Iaccarino.
The second half of this stretch concerns how the simulator represents filaments as a search problem. The authors say that Microcosmos "is highly flexible in terms of the representation of filaments," supporting both direct and indirect encoding schemes. Both schemes map from searchable parameters, which the authors call the genotype, to the physical form and behavior of the creature, which they call the phenotype. The genotype is what an evolutionary search adjusts. The phenotype is the body and movement that results.
In direct encoding, the mapping is one-to-one. Every individual parameter is specified and can be optimized. In Microcosmos, this means the control signals and the physical parameters are set separately for each node. The parameters include a rest angle, a rest length, and bending, shear and stretching stiffness, each defined for every node in the filament.
Indirect encoding works differently. In the authors' description, it "uses a generative or developmental model" that has its own searchable parameters, so the search adjusts the rules that build the body rather than each part of the body directly.
The first experiment checks whether the fluid physics behaves realistically. The authors built five creatures with hand-designed geometry and locomotion: a worm, a tadpole, a ray, a cilia-driven swimmer and a jellyfish. Each was run through a sweep of fluid viscosity, which is how thick the surrounding fluid is, and the total distance each creature travelled was recorded in lattice units, the simulator's own distance measure. The paper's caption presents the result as a confirmation of the engine's "realistic fluid dynamics".
The pattern across the sweep is the key result. The authors tie it to Purcell's scallop theorem, a known result about swimming at very low speeds in thick fluid. In that regime, a body that simply repeats the same shape changes in reverse cannot push itself forward. The caption states that "time-reversible strategies (e.g. ray) produce zero net displacement at high viscosity", while non-reciprocal strategies, such as the cilia, "maintain viable locomotion". The creatures therefore behave in the way the physics predicts. The paper points to animations in its supplementary materials for viewers who want to see the movement.
The paper then turns to how a creature's body is specified. Its encoding schemes generate the physical traits of each node from spatial-temporal or sensory inputs. The examples given are compositional pattern-producing networks, L-systems, which are rule-based growth systems, and cellular automata, which update a grid of cells by local rules. According to the paper, the Microcosmos engine is "search-method agnostic and end-to-end differentiable". That means it does not depend on one way of searching for good designs, and that the simulation can be differentiated, so that gradients can guide optimization.
The experiments use that flexibility in different ways. Filament folding experiments use gradient-based optimization. The neuroevolution experiments, which evolve a controller circuit, and the search for locomotion strategies use quality-diversity search, a method that looks for many different high-performing solutions rather than a single best one. The locomotion experiments with hand-selected parameters serve a narrower purpose: they check that the coupling between the fluid and the filaments works.
The paper frames its validation around four goals, stated in its own words: the engine is validated through experiments "targeting physical correctness, differentiability, evolvability, and scalability." The hand-designed swimmers address the first goal. The later experiments address the others.
The paper's first set of experiments tests whether its simulated physics holds up, using five creatures whose shapes and movement strategies were designed by hand rather than evolved. The authors sweep viscosity, the thickness of the fluid the creatures move through, from one thousandth to two. That range corresponds to a Reynolds number, a measure of how much a swimmer's own inertia matters compared with the fluid's resistance, running from roughly 1,000 down to about 1, depending on the swimmer. At the high end inertia dominates, while at the low end viscosity dominates, the regime in which a microscopic organism pushes through something closer to syrup. The authors state a caveat here: "we acknowledge that outside this range our fluid-filament coupling may give rise to artifacts."
The five creatures are a worm, a tadpole, a ray, a cilia-driven body and a jellyfish. The worm and the tadpole both move with a sinusoidal wave running along a soft filament, the same kind of undulation found in the nematode C. elegans, which the paper discusses earlier. The tadpole is slower because its head creates drag. The ray beats using a single time-dependent angle constraint, a rule that controls how one part of the body is allowed to move over time, and the jellyfish contracts through uniform distance and bending constraints on an open ring. The cilia creature has short filaments attached around a stiff spherical body and moves through asymmetric strokes. The paper says that the peak displacement at high viscosity "is a hallmark of asymmetrical swimming strokes," because propulsion in that regime relies on the greater fluid traction that higher viscous resistance produces. The authors argue that cilia exploit drag asymmetry more directly than the worm's waves do: a stiff stroke sweeps fluid sideways while the recovery stroke curls tightly, which makes them better suited to the very low Reynolds numbers. The passage ends on the limit of the sweep:
Ultimately, a very high viscosity brings all movement close to a stop.
The next experiment tests whether the engine can be trained, not just run. The paper wants to show that a computer can adjust a simulated body directly by following gradients, which are numbers that show how a small change in each parameter would change the result. The authors state the aim plainly: "To demonstrate the differentiability and expressivity of the simulation and filament parameterization, we optimize filaments to fold into target shapes." They describe the task as "a simplified 2D analog of the protein folding problem", a stand-in for the way a chain-like molecule settles into a particular form.
The filament is a chain of 1000 connected nodes, and its angles and segment lengths are the parameters being tuned. The optimizer is Adam, a common gradient-based method, run with a learning rate of 0.04, which sets how large each adjustment is. It takes 300 update steps. Each attempt starts from a random configuration of the filament, which is then simulated for 250 steps. The target is a handwritten digit from the MNIST collection, a widely used set of digit images. The filament's particles are rendered as a density image at 128 by 128 resolution, using a method called Gaussian splatting that spreads each particle into a soft blob. The paper measures the gap between that image and the target digit using mean squared error, a standard score of how far apart two images are.
The key claim is that this error can be traced back through the whole simulation. The paper says that the gradients of the loss "flow end-to-end through the full simulation via JAX automatic differentiation", and that this includes "the PBD constraint solver and fluid computation, directly updating the morphological parameters." JAX is a software library that computes these derivatives automatically. PBD, position-based dynamics, is a common way of keeping the links of a simulated chain at their set lengths while it moves.
The paper reports results for all ten digit classes, zero through nine. For each digit it shows the filament at an early, a middle and a final stage of the folding, beside the target. The authors say that gradient descent "reliably recovers parameters that fold the filament into recognizable digit shapes across all ten classes." They are careful about the quality of the match:
Although the reproductions are imperfect, the results confirm that the Cosserat rod formulation provides a sufficiently expressive and differentiable substrate for 2D morphogenesis.
A Cosserat rod is a physics model of a thin, flexible rod that tracks how each segment bends and twists, and morphogenesis is the process by which living things take on their shapes. So the claim is modest: the model is rich enough to be shaped by gradient descent, even though the digits come out rough.
The researchers also tried a different way of producing movement: evolving the controller instead of designing it by hand. The controller is a Compositional Pattern Producing Network, or CPPN, a small network that maps inputs to outputs and is often used to generate patterns. The paper uses one shared CPPN for every node along a filament. The authors write that they "used one shared CPPN to perform sensorimotor control for each node", so each node senses its local situation and acts on it. Each node receives two inputs: its position along the filament, scaled from 0.0 at one end to 1.0 at the other, and a global timer. The authors say this gives each node "a basic sense of space and time". From these inputs the CPPN produces two outputs for each node, a rest angle and a rest length, which set the shape the node relaxes toward and so let it control its own movement.
The network is not designed by hand but evolved. The paper uses NEAT, short for NeuroEvolution of Augmenting Topologies, a method that repeatedly mutates a network's structure, its activation functions and its connection weights, keeping the variants that perform best. The implementation is built on TensorNEAT. The fitness score is the total displacement the filament travels. The authors report that "the highest fitness swimmers predominantly perform sinusoidal locomotion", meaning the best-performing filaments swim with a smooth, wave-like motion. The paper's supplementary materials include animations of these runs.
The team then added an energy field, described as "a scalar field with scattered packets representing local energy akin to food". A scalar field assigns a single number to every point in space, so the packets mark places where energy can be collected. The CPPN receives the energy value at each node's position as an additional input, and fitness becomes the total displacement plus the total energy collected. The authors say this rewards both movement and chemotaxis, the tendency to move toward a source of food or chemicals, so the evolved filaments are scored for seeking food as well as for swimming.
After the earlier experiments on hand-designed locomotion and neuroevolution, the authors turn to quality-diversity search, a family of methods that looks for many different good solutions instead of a single best one. The paper states its aim plainly: "To discover more diverse types of locomotion, we apply QD search." The specific method is MAP-Elites, which keeps an archive of the best solution found for each cell of a grid, where each cell stands for a different kind of behavior.
The grid is defined by two behavioral descriptors, which are numbers that summarize how a creature moves. The paper uses bending effort and stretching effort. Bending effort adds up, over the course of the simulation, the squared changes in the body's rest angles from one time step to the next. Stretching effort does the same with rest lengths. Rest angles and rest lengths are the shape the body is trying to hold at each moment, so the two measures capture how much a creature keeps reshaping its body as it moves. Because the changes are squared and summed, larger values mean more sustained bending or stretching.
The results come from CPPN-QD search, which is quality-diversity search run with a compositional pattern-producing network. That is a small neural network that generates patterns, and the paper uses it as an encoding of the creature designs. The authors report that
A diverse set of swimming strategies emerged from the CPPN-QD search.
They group these strategies roughly into a few categories. One is sinusoidal movement, a smooth, wave-like back-and-forth motion. Another is directional turning. A third is complex motions that do not fall into the simpler groups. The last is circular structure paddling with "flippers." The supplementary materials include animations of these strategies, so the motions can be seen as well as described.
To test whether the simulator keeps up as crowds grow, the authors measured wall clock time, the real elapsed time a run takes, while increasing the number of particles. They held the grid at a fixed resolution of 256 by 256 cells, ran each simulation for 1000 steps, and went up to 500 thousand particles. The runs used one NVIDIA L40S graphics processor, with JAX version 0.8.1 as the software framework.
The result is stated plainly: "Microcosmos scales linearly." The comparison matters. Many simulators compute how every particle interacts with every other particle, and Particle Life is the example given. Because each particle is checked against the others, the work grows with the square of the particle count, so doubling the crowd roughly quadruples the cost. The authors note that tiling strategies, which group nearby particles to cut down these pairwise checks, still slide back to quadratic growth when particles are unevenly spread out.
The authors tie this result to their design. In their words,
Every design choice in Microcosmos, from grid-based steric repulsion to the local updates of LBM, targets linear scaling.
Steric repulsion is the push that keeps bodies from overlapping, and LBM, the lattice Boltzmann method, is the approach used to simulate the surrounding fluid. Both are built to work only on nearby cells, which is why the cost stays proportional to the number of particles.
The closing discussion opens with the authors' central diagnosis: that artificial life has been held back not by a lack of imagination but by "a lack of the right tools." They argue that existing simulation platforms have forced researchers to choose between "physical credibility and computational tractability." Microcosmos is their attempt to narrow that gap by borrowing from fluid dynamics, biophysics and computer graphics, fields they say the artificial life community has largely not yet absorbed. They also concede that, like any simulation, it "abstracts away much of the physical world." The engine they describe models artificial lifeforms as elastic filament chains moving through a two-dimensional viscous fluid. The authors note that filaments are "among the most ancient and ubiquitous structures in biology," and that their nearest-neighbor structure makes them naturally scalable to simulate. The fluid is solved with the Lattice Boltzmann Method, a numerical technique for simulating how fluids flow. The authors say this imposes real physical constraints on locomotion, which grounds the evolutionary dynamics in physical reality. The whole pipeline is implemented in JAX, a numerical computing library, and is end-to-end differentiable, meaning the effect of every parameter on the outcome can be calculated. The authors say this supports both gradient-based optimization and evolutionary search. According to the authors, their experiments validate the physical correctness of the fluid coupling, the differentiability of the full pipeline, and the ability to automatically discover diverse locomotion and chemotaxis behaviors, the latter meaning movement toward or away from a chemical gradient. They report that all of this scales linearly with particle count. The authors then state the limits plainly. The filament topologies are currently restricted to cactus graphs, a particular family of graph shapes, and the physics engine has a number of hyperparameters, or tunable settings, that may require adjustment. They offer Microcosmos as an open platform and issue a call to the wider community to build on it. Their future work will focus on multi-agent interactions, aiming at larger ecosystems in which unpredictable interactions between individuals can emerge. They believe this will require some notion of birth, death and self-assembly, and that it would open studies of competition and symbiosis, as well as the emergence of individuality and function. Looking further ahead, they hope to see large-scale compute runs applied to open-ended evolution in physically grounded virtual worlds. Their closing analogy is that
Just as scaling compute transformed modern AI, we believe scaling physically grounded simulations may do the same for artificial life.
The paper's main takeaway is that artificial life has been limited less by a lack of ideas than by a lack of suitable tools. Microcosmos offers one attempt at such a tool: a physically grounded, GPU-based, differentiable simulator that can be searched by gradient descent, by evolution and by quality-diversity methods, and that the authors release as an open platform. The experiments support the claims that the fluid coupling respects known physical constraints, that the full pipeline can be trained, that varied swimming strategies can be found automatically, and that cost grows linearly with particle count.
The authors state their limits plainly. Filament topologies are currently restricted to a particular family of graph shapes, the physics engine has tunable settings that may need adjustment, and like any simulation it abstracts away much of the physical world.
The open questions point ahead. The authors plan to study multi-agent interactions in larger ecosystems, which they believe will require some notion of birth, death and self-assembly. They hope large-scale compute runs will be applied to open-ended evolution in physically grounded virtual worlds. Their closing analogy is:
Just as scaling compute transformed modern AI, we believe scaling physically grounded simulations may do the same for artificial life.

