---
title: Self-Organizing Nervous Systems
url: https://www.emergentmind.com/topics/self-organizing-nervous-system-sons
type: topic
---

# Self-Organizing Nervous Systems

Searching arXiv for recent and related papers on Self-Organizing Nervous Systems and adjacent concepts.
Self-Organizing Nervous System (SoNS) denotes a class of nervous-system-like organizations in which coordination structure is established, maintained, and reconfigured by the system itself rather than fixed entirely before operation. In the robotics literature, the term refers most explicitly to a robot-swarm architecture based on a self-organized hierarchy that can merge, split, replace leaders, and coordinate sensing, actuation, and decision-making through a dynamic rooted tree [2401.13103]. In the neuroscience-oriented literature, closely related usage treats the human brain as a self-organizing nervous system because its organization is progressively shaped by activity-dependent interactions that couple stability, plasticity, and adaptive growth [2011.05868]. Related work extends the same general idea toward physically self-assembling robots, developmental growth of artificial neural networks, and multiscale theories of brain dynamics [1505.07050].

## 1. Conceptual scope and defining features

The term SoNS is used across at least two closely connected research traditions. One tradition is architectural and robotic: SoNS is a “novel robot swarm architecture based on self-organized hierarchy,” designed to escape the usual centralized–decentralized dichotomy by allowing robots to autonomously establish, maintain, and reconfigure dynamic multi-level system architectures [2401.13103]. The other tradition is descriptive and theoretical: the nervous system is considered self-organizing when structure and function co-evolve through local interactions, yielding auto-regulation, stability, adaptation to new constraints, and functional autonomy [2011.05868].

In the brain-oriented account, seven properties are identified as characteristic of self-organization in brain systems: modular connectivity, unsupervised learning, adaptive ability, functional resiliency, functional plasticity, from-local-to-global functional organization, and dynamic system growth [2011.05868]. In the swarm-robotics account, four features are emphasized: self-organized controllable hierarchy, interchangeable leadership, explicit inter-system reconfiguration, and reconfigurable swarm behavior structures [2401.13103]. These formulations are not identical, but they converge on a common motif: coordination is distributed in origin, yet it can transiently create organized control structure.

A plausible implication is that SoNS is best understood not as a single algorithm, but as a family of mechanisms for producing nervous-system-like control from local interactions. In some realizations this takes the form of a temporary “brain” within a robot collective; in others it denotes self-maintaining neural organization without any externally imposed blueprint.

## 2. Swarm-robotics realization: hierarchy, information flow, and control

The canonical robotic SoNS is modeled as a directed rooted tree \(G=\{V,E\}\), with robots as vertices, directed parent→child links as edges, and a root node \(r_1\) serving as the SoNS-brain [2401.13103]. The hierarchy is formed through temporary, bidirectional parent–child links. A child sends sensor information upstream; a parent sends actuation instructions downstream. Information therefore propagates hop-by-hop through the hierarchy rather than by direct brain-to-every-robot communication.

The architecture begins with each robot as a single-robot SoNS. A brain robot holds a target map containing target topology \(G\) and target node/link attributes \(A\). Robots search for peers, exchange recruitment messages, agree which becomes parent and which becomes child, and recursively pass target subgraphs to descendants until the target structure is formed. The supplementary formalization describes a target graph \(G=(V,E,D)\), where \(D\) stores desired formation attributes such as relative positions, and each robot knows only its own local state, the attributes of its immediate neighbors, and its target subgraph \(H_{r_i}^{*}\), not the full global graph [2401.13103].

Several node attributes support self-organization. Each robot stores \(SoNSrootID_i\) and \(SoNSrootRANK_i\), propagated downstream from the brain. Subtree size is represented by downstream vertex cardinality \(\lambda^{card}(r_i)\) and type-specific versions \(\lambda^{card}_{\tau_\ell}(r_i)\). Depth-related structure is summarized by vertex height:
$$
\lambda^{high}(r_i) \leftarrow \max(\lambda^{high}(r_j) \mid r_j \in C_{r_i}) + 1
$$
These quantities are used for recruitment ordering, reconfiguration timing, and avoiding stale SoNS identities after merge–split events [2401.13103].

Control is decomposed into local, hierarchical, and global components. The motion outputs are
$$
\boldsymbol{v}_{r_i}^{*} = \boldsymbol{v}_{r_i}^{hierarchical} + \boldsymbol{v}_{r_i}^{local} + \boldsymbol{v}_{r_i}^{global}
$$
and
$$
\boldsymbol{\omega}_{r_i}^{*} = \boldsymbol{\omega}_{r_i}^{hierarchical} + \boldsymbol{\omega}_{r_i}^{local} + \boldsymbol{\omega}_{r_i}^{global}.
$$
This decomposition allows local obstacle avoidance, parent-propagated mission references, and SoNS-wide goals to be superposed within one control interface [2401.13103].

The same work provides a leader–follower analysis in which, for a follower \(R_j\) relative to leader \(R_i\),
$$
\dot{\mathbf{p}_i} = \mathbf{u}_i,\qquad \mathbf{e}_{ij} = \mathbf{z}_{ij} - \mathbf{d}_{ij},
$$
with reactive control law
$$
\mathbf{u}_j = \mathbf{K}^j(\mathbf{z}_{ij} - \mathbf{d}_{ij}) = \mathbf{K}^j \mathbf{e}_{ij}.
$$
The reported results include exponential convergence of formation tracking error for stationary leaders, input-to-state stability when leaders move, and cascaded ISS composition for multi-level hierarchies [2401.13103].

## 3. Reconfiguration, resilience, and online reprogramming

A defining property of SoNS is explicit runtime reconfiguration. Two SoNSs can merge when their brains compare quality; the lower-quality brain becomes a child of the higher-quality brain, with equal-quality cases resolved randomly. A SoNS can split when a parent expels a child, and the expelled robot immediately becomes the brain of a new single-robot SoNS. Parents can hand over children, children can be reassigned to better matching target roles, and links can be broken and re-established while the mission continues [2401.13103].

Role assignment is supported by a node-allocation procedure based on source displacements, source subtree sizes, target displacements, and target subtree sizes. The method constructs cost matrices, builds a flow graph \(G_{flow}\), solves a maximum-flow / assignment problem, matches robots to target roles, and reassigns children if a better match appears. This makes hierarchy itself an object of self-organization rather than a fixed design choice [2401.13103].

Fault tolerance is treated as a first-class architectural capability. Demonstrated cases include single robot replacement, automatic brain replacement, arbitrary permanent failures under high-loss conditions, and recovery from temporary vision or communication failure. In real experiments, however, missions could not continue if all robots of a required type were lost; for example, if both aerial robots failed, the remaining ground robots could not complete the mission requirements [2401.13103].

A later extension positions SoNS as middleware for online automatic code generation in robot swarms [2510.04774]. In that formulation, robots dynamically organize into a rooted communication graph with a designated current SoNS brain. Local sensor information is aggregated and forwarded upstream; global actuation commands and updated code flow downstream. The paper distinguishes local actuation variables \(\boldsymbol{v}^{Local}\), \(\boldsymbol{w}^{Local}\) from global actuation variables \(\boldsymbol{v}^{Global}\), \(\boldsymbol{w}^{Global}\), and states that the maximum update delay is \(n+1\) steps, where \(n\) is the depth of the rooted graph [2510.04774].

This middleware role is used to support external LLM-assisted repair of swarm behavior. In a demonstration with 6 real robots and simulation trials with 34 robots total over 20 repetitions, the swarm began with no obstacle-avoidance code; when it got stuck, the current SoNS brain sent contextual information to DeepSeek-R1 through the OpenRouter API, received Lua code, updated the mission program, and disseminated it through the SoNS. The reported mission success rate was 85%, corresponding to 17 out of 20 trials [2510.04774]. The open-ended request for assistance, delayed global estimation, and dependence on a current brain robot define the operational limits of this approach as presently described.

## 4. Related robotic nervous systems and developmental analogues

A closely related but distinct architecture is the Virtual Nervous System (VNS) for physically self-assembling robots [1505.07050]. There, a robot’s nervous system is defined as “the processing unit responsible for controlling the robot body, together with the links between the processing unit and the sensorimotor hardware of the robot.” When robots physically connect, their nervous systems merge into a virtual nervous system spanning the connected bodies. The physical connection topology is imposed as a hierarchical tree, the logical topology follows the physical topology, and the unique root becomes the brain; the brain of a newly connecting robot body must cede authority to that root [1505.07050].

VNS-based robots can merge into larger bodies with one brain, split into separate bodies with independent brains, use sensing and actuating capabilities of specialized peers, and self-heal by removing or replacing malfunctioning body parts, including the brain. Recursive self-knowledge of descendants allows rapid merging and splitting without expensive global rediscovery. In a ten-robot experiment, robots self-assembled into two composite robots, each with one brain, and later merged into one larger composite with a single brain [1505.07050]. The same work is explicit that current behavior is pre-programmed: desired body configurations are specified in advance rather than autonomously selected.

Self-organizing nervous-system ideas also appear in neural developmental programs for artificial neural networks [2307.08197]. The Neural Developmental Program (NDP) is a distributed neural controller that lives inside each node or cell of a growing network and operates through local communication alone. Starting from a seed graph, the developmental process repeatedly computes graph diameter, propagates node states through graph convolution or message passing, selects growing nodes, adds new nodes, initializes new embeddings as the mean of parent embeddings, predicts edge weights, and optionally prunes edges below a threshold \(\mathcal{P}\) [2307.08197]. The genotype therefore remains constant-size even as the phenotype grows.

The reported results establish feasibility rather than a final solution. In the evolutionary setting, the NDP solves XOR with a 4-node, 7-edge graph, solves CartPole with a 10-node, 33-edge policy achieving \(500 \pm 0\) reward, nearly solves LunarLander with \(116 \pm 124\) reward, and grows a graph with small-world coefficients satisfying the criterion. In the differentiable setting, it achieves 91% on 8×8 MNIST and 29 reward on HalfCheetah with behavior cloning [2307.08197]. The authors identify missing ingredients including activity-dependent and reward-modulated development, learned stopping criteria, and scaling to richer environments.

An earlier self-organized agent-based simulation model for exploration of human synaptic connections provides another analogue [1207.3760]. That framework uses Neuron agents, Muscle agents, Axons, a WiringViewer observer, Non-Cooperative Situations such as DepolarizationNCS and InstantFrequencyNCS, and AdaptiveTrackers for parameter tuning. Its aim is to estimate unknown synaptic connections by comparing simulated outputs with reference data and reorganizing the agent network until functional adequacy is improved [1207.3760]. The contribution is preliminary and conceptual, but it exemplifies SoNS-style inference through local feedback, reorganization behavior, and evolution behavior.

## 5. Brain self-organization, criticality, and multiscale dynamics

In the neurobiological account, the human brain is presented as a paradigmatic self-organizing nervous system because its functional organization is generated, maintained, and updated through activity-dependent interactions among neurons, circuits, and large-scale networks [2011.05868]. Modular connectivity reduces complexity; unsupervised learning is grounded in Hebbian reinforcement, Long-Term Potentiation, and Long-Term Depression; adaptive ability is linked to synaptic plasticity and neuromodulatory control; functional resiliency depends especially on indirect, long-range connections; functional plasticity allows one subsystem to take over the function of another; and from-local-to-global functional organization explains how coherent large-scale function emerges from local interaction [2011.05868]. Dynamic system growth is defined functionally rather than merely anatomically: a fixed-size network can generate new functions through activity-dependent reorganization and resonant circuitry formation.

This line of argument is extended by work on fractals, power-law scaling, and self-organized criticality in the nervous system [0910.2741]. Fractality is described at morphological, cellular, mesoscopic, and macroscopic levels. The review emphasizes that fractality, self-similarity, power-law scaling, and self-organized criticality are related but not identical, and that power-law scaling alone is not proof of self-organized criticality. At the mesoscopic level, the review reports neuronal avalanches with size and lifetime distributions following power laws with exponent approximately \(-3/2\), and a branching parameter near \(\sigma = 1.04\), close to the critical value 1 [0910.2741]. The same account proposes adaptive advantages including economical wiring, efficient transport, long-memory adaptation, rapid state switching, and integration across multiple levels of organization.

Theoretical caution is developed further in work on self-organized bistability (SOB) and cortical dynamics [1911.05382]. There, the fast activity field \(\rho\) and slow field \(E\) in SOB,
$$
\dot{\rho}(\vec{x},t) = [E(\vec{x},t)-a]\rho + b\rho^{2}-\rho^{3}+D\nabla^{2}\rho + h + \eta(\vec{x},t),
$$
$$
\dot{E}(\vec{x},t) = \nabla^{2}\rho(\vec{x},t)-\varepsilon \rho(\vec{x},t)+h,
$$
are compared with a cortical Landau–Ginzburg model using synaptic resources \(R\):
$$
\dot{\rho}(\vec x, t)=(R(\vec x, t)-a)\rho+b\rho^2-\rho^{3}+I + D\nabla^{2}\rho+ \eta(\vec x, t),
$$
$$
\dot{R}(\vec x, t)=\frac{1}{\tau_R}(\xi-R)- \frac{1}{\tau_D} R\rho.
$$
The conclusion is qualified: in the limit of infinitely slow synaptic dynamics the two theories become identical, but the timescales required for exact self-organization do not seem to be biologically plausible. For realistic regimes, the authors introduce the notion of “self-organized collective oscillations” rather than ideal SOB [1911.05382].

A different theoretical direction models the nervous system as an open quantum system [2104.09424]. Neurons are treated as a system of two-level units, while dendrites, axons, synapses, glial cells, and surrounding circuit-like electromagnetic modes form an environment of oscillators. Effective neuron–neuron couplings then emerge through the environment rather than being inserted directly as a fixed synaptic graph. The reduced dynamics are described by a Born–Markov master equation, and in a collective uniform-coupling limit the model yields stimulus-responsive, action-potential-like collective behavior [2104.09424]. This does not define SoNS in the robotic sense, but it is consistent with the broader idea that nervous-system-level coordination can emerge from local interactions mediated by shared dynamical substrates.

## 6. Empirical demonstrations, limitations, and open questions

The robotic SoNS literature reports proof-of-concept missions including establishing self-organized hierarchy, balancing global and local goals, collective sensing and actuation, binary decision-making, and splitting-and-merging search-and-rescue in real heterogeneous aerial-ground swarms [2401.13103]. Real experiments use e-puck ground robots and custom S-drone quadrotors, while simulator studies extend to 250 robots, specifically 200 ground robots and 50 aerial robots. For establishment missions, the measured quantities include actuation error, communication load in bytes, computation load in CPU cycles, and convergence time. The reported scalability pattern is that performance is fully reliable up to 125 robots, somewhat stable but with occasional non-convergence between 125 and 220 robots, and substantially degraded above 220 robots; steady-state actuation error remains low, under about 0.5 m [2401.13103].

These results establish SoNS as a workable architecture, but the limitations are explicit. The main experiments use a simple reactive controller, not a feedforward one; real hardware depends on vision-based relative sensing; aerial hover noise contributes noticeable error; local communication depends on visibility constraints; and scalability degrades above roughly 220 robots in the current implementation [2401.13103]. In the LLM-mediated extension, the task structure is intentionally simple, safety requirements are basic, and some failed trials occurred because generated code removed mechanisms needed for monitoring whether robots were stuck [2510.04774].

In related systems, limitations are equally clear. VNS currently relies on pre-programmed desired body configurations rather than autonomous choice of morphology [1505.07050]. NDP growth is not activity-dependent during an agent’s lifetime, performance can degrade after too many growth steps, and the NDP itself can be larger than the policy network it produces [2307.08197]. In cortical theory, exact SOB requires implausibly slow synaptic timescales, which weakens strict self-organization claims for short-term plasticity as the operative mechanism [1911.05382]. In the fractality literature, power-law scaling is repeatedly distinguished from proof of criticality, and the relation among fractals, self-similarity, and self-organized criticality remains unresolved [0910.2741].

Taken together, these constraints delimit the present state of the field. SoNS is not a single settled doctrine. It is a convergent research framework spanning robot swarms, modular self-assembly, neural developmental growth, and theories of brain organization. Across these settings, the recurring technical claim is that useful nervous-system-like coordination can arise from local interactions, recursive structure, and dynamic reconfiguration; the recurring unresolved issue is how far such self-organization can be pushed without reintroducing hidden centralization, brittle assumptions, or biologically implausible idealizations.

Source: https://www.emergentmind.com/topics/self-organizing-nervous-system-sons