SwarmEvolve: Evolving Decentralized Swarm Behaviors

• SwarmEvolve is an evolutionary-swarm framework for synthesizing decentralized controllers that optimize swarm behavior through local sensing and actuation, designed to achieve diverse collective actions.
• SwarmEvolve uses frameworks like genetic algorithms, neural networks, and Markov models to generate adaptive behaviors, applicable to robotics and AI where collective action is essential.
• Edge on technologies that utilize SwarmEvolve are Commercial robotics, autonomous vehicles, and AI-driven traffic management.

SwarmEvolve is an evolutionary-swarm framework concept for synthesizing decentralized controllers whose performance is evaluated at the collective level. Its principal design problem is the transformation of local sensing and actuation rules into emergent macroscopic behavior: aggregation, dispersion, navigation, foraging, communication, monitoring, predator avoidance, environmental adaptation, or open-ended collective organization. The term encompasses several complementary methodological traditions, including neuroevolution, novelty search, quality-diversity algorithms, grammatical evolution, reinforcement learning, stigmergic communication, coevolution, lifelong evolution, and mean-field system identification. Across these approaches, the evolutionary individual may be a shared controller, a heterogeneous collection of agent controllers, a behavioral archive, or an emergent collective structure.

1. Conceptual foundations and scope

SwarmEvolve addresses the genotype-to-collective-behavior chain

$$\text{genotype}\rightarrow\text{controller phenotype}\rightarrow\text{individual behavior}\rightarrow\text{group behavior}\rightarrow\text{fitness}.$$

The defining characteristic is that local controllers are evaluated through their effects on a swarm rather than solely through isolated-agent performance. A single genome may be copied across all robots, as in evolutionary gradient following, aquatic-surface-robot control, and predator–prey reinforcement learning. Alternatively, each agent may possess a distinct rule set, as in decentralized grammatical evolution and quality-diversity methods.

The framework has several recurring properties:

• Decentralized execution: agents use local observations and do not rely on a centralized runtime controller.
• Collective evaluation: fitness depends on swarm-level outcomes such as food returned, area monitored, packets delivered, survival, or environmental-field value.
• Emergent coordination: alignment, flocking, stigmergy, role differentiation, and collective navigation may arise without being explicit objectives.
• Embodied interaction: agents affect one another through physical contact, local sensing, communication, shared resources, or environmental traces.
• Evolutionary variation: parameters, topologies, rule sets, communication mechanisms, or controller repertoires are modified through evolutionary operators.
• Robustness assessment: controllers are tested across swarm sizes, randomized initial states, changing environments, faults, noise, or unseen layouts.

This formulation differs from conventional single-agent optimization because the fitness landscape is generated partly by interactions among concurrently executing agents. It also differs from manually designed swarm control because the microscopic rules need not be specified in advance. Evolutionary robotics studies demonstrated this pattern for aggregation and energy-sharing swarms, where novelty search could discover collective stepping stones that objective-based evolution often suppressed (Gomes et al., 2013). More recent work applied the same principle to environmental gradient following, where a single evolved controller produced collective motion despite receiving only local environmental and neighbor information (Diggelen et al., 2022).

2. Controller representations and evolutionary substrates

Neural and topology-evolving controllers

NEAT is a recurrent representation used in several SwarmEvolve-relevant systems. It evolves connection weights and network topology while protecting structural innovations through historical markings, speciation, and fitness sharing. In simulated homogeneous swarms, NEAT controllers were evolved for aggregation and charging-station sharing using population size $200$, $250$ or $400$ generations, recurrent links, and repeated stochastic evaluation (Gomes et al., 2013).

A real aquatic-surface-robot system used NEAT to evolve shared neural controllers for homing, dispersion, clustering, and area monitoring. The controller received local waypoint, robot, and geofence observations and generated linear-speed and angular-velocity commands. Controllers evolved in simulation were transferred to up to ten physical robots operating in an uncontrolled outdoor water body, with local Wi-Fi communication and no centralized execution controller (Duarte et al., 2015).

Fixed or partially evolved neural architectures are also used when the evolutionary search cost is high. In environmental gradient following, two hidden reservoir layers were fixed randomly and only an 18-dimensional output-weight vector was evolved with Differential Evolution. The robots received local scalar-field values and nearby-neighbor geometry; collective gradient following emerged despite the absence of explicit communication or memory (Diggelen et al., 2022).

Spiking neural networks provide another substrate. In an ant-foraging model, a shared network with 12 input neurons, 20 middle-layer neurons, and four output neurons controlled movement and pheromone deposition. Synaptic weights and delays were evolved with a genetic algorithm. The pheromone-enabled colony returned a median of $137$ out of $150$ food units, compared with $117.5$ for a rule-based baseline and $97$ for a colony evolved without pheromone sensing (Romero et al., 2022).

Markov Networks and Markov Brains

Probabilistic Markov Networks and Markov Brains support compact controllers with memory and stochastic or deterministic state transitions. A predator–prey model used Markov Networks with 24 sensory inputs, two actuator outputs, and 38 hidden states. Local binary vision and movement rules were sufficient for predator-dependent evolution of cohesive or dispersed prey groups (Olson et al., 2013).

A separate study evolved Markov Brains in swarms of $1$, $18$, $36$, $54$, or $72$ animats. The task involved repeated movement between two rooms while avoiding collisions. Controllers evolved at an intermediate swarm size of $36$ displayed the strongest cross-size generalization, behavioral flexibility, and largest strongly connected components, suggesting that evolutionary swarm size is itself a critical design variable (Fischer et al., 2018).

Grammatical evolution and behavior trees

SLCS2 evolves heterogeneous condition–action rule sets using a grammar, online Q-learning, novelty search, and partial rule sharing. Each agent receives a rule set of $100$ rules, with conditions based on packet state, neighbors, obstacles, distances, interference, and network state. High-contribution agents donate strong rules to low-contribution agents, while low-prediction-error recipient rules are retained (Smith et al., 2019).

BeTr-GEESE combines Grammatical Evolution with PPA Behavior Trees. Each genome is a 100-bit string mapped through a grammar into a behavior tree with postconditions, preconditions, constraints, and actions. Behavior-tree statuses provide task-independent feedback. The method achieved a reported $75\%$ success rate in evolving foraging and nest-maintenance behaviors, compared with $9.3\%$ for GEESE-BT (Neupane et al., 2022). Its principal limitation is that performance depends on continued evolution: static controllers generally contain only partial subtasks.

Population-based parameter search

Evolution strategies can operate directly on controller parameters. A visual-navigation framework used a pretrained visual encoder and a compact approximately $3{,}704$-parameter MLP, then optimized the shared controller with the Cross-Entropy Method or CMA-ES. CEM achieved $20.38$ mean explored sectors, while CMA-ES achieved $19.44$; the CEM policy was statistically comparable to a LiDAR controller in sector coverage and reduced reported average energy consumption relative to the LiDAR method (Díez et al., 5 Oct 2026).

EM-PSO provides another derivative-free parameter optimizer, using exponentially averaged velocity memory:

$M_i^{d+1}=\beta M_i^d+(1-\beta)v_i^d,$

followed by personal-best and global-best attraction. Its application to a neural Schrödinger solver used $\beta=0.9$, $c_1=0.8$, and $c_2=0.9$, but the evidence was limited to a low-dimensional particle-in-a-box benchmark and did not establish superiority over gradient-based optimization (Jivani et al., 2021).

3. Selection pressure, diversity, and behavioral archives

Fitness-based evolution and deception

Fitness-based evolution is effective when the objective provides an informative gradient, but it can converge prematurely when collective success requires poorly rewarded intermediate behaviors. In the charging-station task, remaining stationary conserved energy and one robot occupying the station could outperform active search during early generations, even though neither behavior solved the collective survival problem (Gomes et al., 2013).

Novelty search replaces immediate task quality with behavioral difference from previously observed behaviors. A behavioral descriptor $b$ is compared using a distance such as

$dist(x,y)=\sqrt{\sum_{j=1}^{D}(x_j-y_j)^2},$

and novelty is estimated from the average distance to the $k$ nearest neighbors in the current population and archive:

$\rho(x)=\frac{1}{k}\sum_{i=1}^{k}dist(x,\mu_i).$

In the swarm experiments, the archive used stochastic insertion probability $0.02$ and $k=15$. Novelty search produced comparable aggregation performance to fitness evolution and substantially more reliable charging-station solutions. It also discovered lower-complexity controllers and diverse strategies such as wall following, flocking, circular exploration, and dynamic cluster formation (Gomes et al., 2013).

Descriptor design is consequential. Macroscopic descriptors based on center-of-mass distance and cluster count generated different behavioral repertoires. Adding weakly task-related dimensions could harm novelty search by directing exploration toward easy but unproductive regions. Hybrid methods, including progressive minimal-criteria novelty search and linear novelty–fitness scalarization, improved performance when descriptors were broad or partially misaligned with task success.

Quality diversity and decentralized archives

EDQD—Embodied Distributed Quality Diversity—combines MAP-Elites with the decentralized evolutionary algorithm mEDEA. Each robot maintains a local archive indexed by two functional traits: maximum distance from its starting point and red-to-blue token collection ratio. The archive contains $15\times15=225$ cells, with one elite per behavioral niche (Hart et al., 2018).

EDQD variants differ in archive transmission:

• EDQD-R randomly selects one received archive.
• EDQD-M1 merges received archives.
• EDQD-M2 merges received archives with a persistent MemoryMap.
• EDQD-M3 merges received archives into the robot’s LocalMap and rebroadcasts inherited information.

The median number of occupied cells was $87$, $94$, $91$, and $93$ for EDQD-R, M1, M2, and M3, respectively, compared with $69$ for $mEDEA_{fps}$. EDQD-M3 significantly improved local-map precision, while all EDQD variants produced substantially more diverse final swarms than the baseline. The mechanism maintains diversity through niche-indexed elitism rather than geographical reproductive isolation.

Open-endedness and emergent objects

Swarm Chemistry explores a more radical architecture in which local particles transmit and mutate behavioral recipes through collisions, while macroscopic objects are detected retrospectively. Long simulations with $10{,}000$ particles, $300{,}000$ time steps, and spatial perturbations produced $49{,}540$ harvested objects (Sayama, 2018).

The study distinguishes microscopic novelty from macroscopic organization. Evolutionary exploration was approximated by newly appearing colors, while macroscopic structuredness was measured using the Kullback–Leibler divergence between observed and random pairwise-distance distributions. The system continued producing microscopic variants but generally entered stable macroscopic attractors after approximately $100{,}000$ steps. This indicates autonomous creativity without demonstrating unrestricted open-ended evolution.

A related theoretical framework argues that communication, concurrency, and complexity or population size can generate new collective regimes. Its proposed mechanism is exploration of the adjacent possible through constrained communication and multiple timescales (Witkowski et al., 2019). These claims remain architectural hypotheses rather than demonstrations of unbounded innovation.

4. Communication, adaptation, and coevolution

Stigmergic communication

Pheromone-mediated communication provides an external memory and a collective credit-assignment mechanism. In the evolved ant-foraging model, ants could deposit and sense pheromone, but no pheromone-following rule was prescribed. Evolution produced spatial trails linking food piles and the nest. Disabling pheromone sensing after evolution reduced the median food return to $0$, demonstrating that the communicating colony had co-adapted its movement policy to the environmental signal (Romero et al., 2022).

A theoretical model of C. elegans foraging formalizes pheromone as a distributed reinforcement signal. If patch $i$ has intrinsic attractiveness $A_i$ and pheromone level $\tau_i(t)$, the collective selection probability is

$$P_i(t)=\frac{\tau_i(t)A_i}{\sum_{j=1}^{M}\tau_j(t)A_j}.$$

Pheromone dynamics are

$\tau_i(t+1)=\rho\tau_i(t)+\Delta_i(t),$

where $\rho$ controls persistence and $\Delta_i(t)$ represents deposition. Under the model’s assumptions, the resulting probability update has the same structure as cross-learning: selected alternatives increase toward probability one, while unselected alternatives decline proportionally to their existing probability (Vellinger et al., 24 Sep 2025).

Persistent stigmergic memory can also cause lock-in. In a dynamic multi-armed-bandit model, homogeneous pheromone-sensitive populations switched to the newly optimal arm in only $13\%$ of runs. Introducing a $10\%$ minority of pheromone-insensitive explorers increased reported success to $100\%$. The result motivates explicit heterogeneity as an exploration reservoir.

Individual learning and inherited adaptation

A distributed evolutionary swarm study compared fixed learning, inherited parameters, and inherited parameters combined with lifetime adaptation. The environment contained positive and negative token types whose values reversed seasonally. EVO+IL, which evolved initial token preferences, learning rate, and learning sign while allowing lifetime updates, was generally the most robust mechanism, particularly when token feedback was abundant (Steyven et al., 2018).

The core design pattern is

$\text{inherited prior}+\text{lifetime correction}.$

Evolution alone performed well in static environments but became maladaptive after seasonal reversal. Fixed individual learning was generally ineffective or harmful. The number of informative experiences mattered more than the magnitude of each reward: many small learning opportunities were more useful than a few large ones.

Predator–prey coevolution

Predator–prey models show how survival pressure can generate aggregation, dispersion, and predator tactics. One model varied predator attack modes and found that peripheral or persistent attacks strongly favored cohesive swarming, while high-density attacks favored dispersion. Density-dependent predation, approximating Hamiltonian domains of danger, allowed selfish-herd behavior to evolve even with coevolving predators (Olson et al., 2013).

A reinforcement-learning predator–prey system used shared actor and critic networks within each species and rewards of $-1$ for prey contact and $+1$ for predator contact, with movement and boundary costs. Prey developed cohesion and local alignment, while predators displayed dispersion tactics, edge predation, and apparent confusion. Reported degree of sparsity decreased from roughly $0.22$ to $0.19$, while degree of alignment increased from about $0.65$ to $0.82$ (Li et al., 2023).

These results require qualification. The reinforcement-learning study uses shared policies and adaptive episodes rather than genetic reproduction, mutation, or inheritance. “Coevolution” therefore refers to reciprocal policy adaptation rather than a complete genetic evolutionary process.

Lifelong evolution

Lifelong Evolution of Swarms studies task sequences such as red-box, green-box, and red-box foraging. A population evolved with NEAT retained previous-task specialists even when the current elite catastrophically forgot them (Leuzzi et al., 22 Mar 2025).

Population-level retention and elite retention are distinct:

$R^{pop}_{t-1}(P_t)=\max_{\pi\in P_t}f_{t-1}(\pi),$

whereas top-individual retention evaluates the controller currently best on the new task. In the reported red–green–red sequence, population-level red retention was approximately $14.64$, while top-individual red retention was approximately $2.08$. Genetic-distance regularization reduced forgetting but imposed a stability–plasticity trade-off. Model-specific regularization achieved green fitness $28.50$, red retention $19.52$, and red forgetting $5.02$.

The population, species structure, and archive thus function as distributed behavioral memory. A SwarmEvolve deployment that retains only one elite controller risks losing capabilities that remain present elsewhere in the population.

5. Embodiment, scalability, and sim-to-real transfer

SwarmEvolve systems are highly sensitive to embodiment. Robot morphology, sensing range, arena geometry, communication topology, environmental physics, and swarm size alter the selective pressures under which controllers evolve.

The aquatic-surface-robot study transferred simulated controllers to real robots by combining empirical motion calibration, shared middleware, randomized initial conditions, and conservative noise. Homing and monitoring transferred successfully, while dispersion exposed a motion-model mismatch: real robots executed larger circles and drifted away from their target positions (Duarte et al., 2015). The study also demonstrated operation under wind, currents, GPS errors, motor faults, battery variation, and heterogeneous maximum speeds, although systematic fault injection was not performed.

Swarm size during evolution can determine generalization. Animats evolved alone did not develop useful social responses, while animats evolved in very large groups became specialized to congestion. The intermediate $36$-agent condition produced the highest cross-size robustness and largest recurrent brain components (Fischer et al., 2018). Similarly, gradient-following controllers showed a practical scalability “sweet spot” at $14$ robots, with lower mean performance for $5$ and $50$ robots under the tested conditions (Diggelen et al., 2022).

Vision-based evolution demonstrates a different scaling strategy. A frozen visual encoder compresses camera images into 32-dimensional latent vectors, allowing CEM or CMA-ES to optimize only a compact shared MLP. The CEM controller achieved comparable exploration to a LiDAR baseline while using less reported sensing and motion energy (Díez et al., 5 Oct 2026).

Mean-field modeling offers another route to large-scale simulation and inference. A Cucker–Smale swarm can be represented microscopically by pairwise alignment dynamics or macroscopically through density and momentum equations. A screened-Poisson interaction kernel with parameters $(k,\lambda)$ was learned from density evolution using PDE-constrained optimization. In one dimension, parameters $(4,1)$ were recovered approximately as $(3.98721701,0.98546559)$; in two dimensions, the recovered values were $(4.01514,1.00194)$ (Mavridis et al., 2021).

The mean-field approach is computationally useful but limited. Density-only observations may identify behaviorally equivalent kernels without uniquely identifying microscopic interactions. The model assumes homogeneous agents, symmetric interaction laws, and a pressureless or monokinetic closure. It does not directly model heterogeneous roles, obstacles, communication delays, or adversarial agents.

6. Applications, limitations, and methodological assessment

SwarmEvolve methods have been applied to:

• aggregation and charging-station sharing;
• homing, dispersion, clustering, and environmental monitoring;
• token gathering and functional diversity;
• predator avoidance and selfish-herd formation;
• multi-hop data transfer;
• ant-style foraging;
• environmental gradient following;
• open-ended object formation;
• visual indoor exploration;
• lifelong adaptation to recurring tasks;
• collective motion inference from density observations.

The applications reveal several recurring trade-offs.

Exploration versus exploitation: Pure novelty search avoids deceptive convergence but may explore irrelevant regions. Hybrid novelty–fitness methods and archive-based quality diversity concentrate search without eliminating behavioral alternatives (Gomes et al., 2013).

Diversity versus current fitness: MAP-Elites preserves lower-fitness behaviors when they occupy distinct functional niches. Such behaviors may become valuable after environmental change (Hart et al., 2018).

Memory versus plasticity: Persistent pheromone, archived controllers, and strong genetic regularization preserve useful information but can obstruct adaptation to new conditions (Vellinger et al., 24 Sep 2025, Leuzzi et al., 22 Mar 2025).

Complexity versus tractability: Richer grammars, larger neural networks, more communication channels, and larger populations increase representational capacity but also expand the search space and evaluation cost (Smith et al., 2019, Neupane et al., 2022).

Task fitness versus emergent structure: A swarm-level objective can produce alignment, flocking, pheromone trails, or collective navigation without explicit terms for those behaviors. However, descriptive emergence does not establish causal necessity, and the observed organization may depend strongly on the task, simulator, reward, and observation model (Diggelen et al., 2022, Li et al., 2023).

Simulation versus physical deployment: Simulation-based evolution supports large evaluation budgets but is vulnerable to the reality gap. Empirical calibration, shared software, randomized noise, hardware-in-the-loop testing, and explicit fault models are required for reliable transfer (Duarte et al., 2015).

Several limitations recur across the literature. Many studies use homogeneous simulated agents, small swarms, task-specific environments, incomplete statistical reporting, or manually selected behavioral descriptors. Some studies describe policy adaptation as evolution without genetic reproduction or inheritance. Others report emergent communication or open-endedness without formal causal or long-term innovation metrics. Claims should therefore distinguish:

• genetic evolution from reinforcement-learning adaptation;
• population-level memory from elite-level memory;
• microscopic novelty from macroscopic innovation;
• behavioral emergence from causal demonstration;
• simulation feasibility from physical robustness;
• task generalization from transfer to genuinely novel environments.

A robust SwarmEvolve implementation would combine a shared or modular local controller, repeated stochastic collective evaluation, behavioral archives, elite preservation, novelty or quality-diversity selection, task-indexed memory, adaptive exploration, and explicit robustness testing. It should log agent trajectories, interaction graphs, communication traces, controller complexity, lineage structure, swarm-level descriptors, and performance across swarm sizes and environmental conditions. For dynamic tasks, population diversity should be preserved as a behavioral resource rather than treating the current best controller as the complete state of knowledge.

The resulting architecture can be summarized as

$$\text{local controller} +\text{collective evaluation} +\text{behavioral diversity} +\text{environmental or social memory} +\text{adaptive evolutionary selection}.$$

SwarmEvolve is therefore best understood not as one fixed algorithm, but as a family of evolutionary-swarm methodologies that treat collective behavior, distributed memory, embodied interaction, and controller diversity as first-class components of the evolutionary process.