Swarm: Distributed Collective Systems

• Swarm is a term for distributed systems where local interactions among simple agents yield emergent global behavior without centralized control.
• Applications range from coordinated UAV, spacecraft, and ground robot teams to decentralized computing architectures handling load balancing and memory replication.
• Key models include macroscopic density–performance, field-based programming, and leaderless protocols that ensure robust, scalable, and fault-tolerant operations.

Searching arXiv for the supplied swarm-related papers to ground the article in current indexing metadata. arxiv_search.query({"^^^^^^^^2search_query2 OR ti:\2^^^^^^^^"SwarMS: Swarm-Computations for Mobile Scenarios\"","max_results":5,"sort_by":"relevance","sort_order":"descending"}) arxiv_search({"query":"SwarMS: Swarm-Computations for Mobile Scenarios","max_results":5}) search_arxiv({"query":"SwarMS: Swarm-Computations for Mobile Scenarios","max_results":5}) Swarm denotes a family of distributed collective systems in which global behavior emerges from local interactions among many relatively simple agents, typically without a central controller or elected leader. In contemporary technical usage, the term spans biological models of collective intelligence, engineered robot teams such as UAVs, spacecraft, and ground robots, and a set of named computing systems that apply swarm-style decentralization to networking, storage, load balancing, and memory replication (&&&^^^^2id:(Scherb et al., 2021) OR ti:\2^^^^&&&, Majee et al., 2024, Chamanbaz et al., 2017, Daghistani et al., 2020).

^^^^2id:(Scherb et al., 2021) OR ti:\2^^^^. Foundational meaning and macroscopic models

Swarm intelligence is defined as the emergent collective intelligence arising from the local interactions of a large number of relatively simple and autonomous agents. Canonical biological inspirations include ants, termites, bees, flocks of birds, schools of fish, bacterial colonies, herds of ungulates, and human crowd dynamics. Across these examples, the recurrent principles are self-organization, stigmergy, positive and negative feedback, and an exploration–exploitation balance regulated by parameters such as pheromone decay rate PRESERVED_PLACEHOLDER_^^^^2search_query2^^^^ or inertia weight PRESERVED_PLACEHOLDER_^^^^2id:(Scherb et al., 2021) OR ti:\2^^^^ (&&&^^^^2id:(Scherb et al., 2021) OR ti:\2^^^^&&&).

In robotic and cyber-physical settings, the same idea is formalized in state-space terms. A UAV swarm may be written as $N$ agents with states $x_i(t)\in\mathbb{R}^n$ evolving under $\dot x_i=f_i(x_i,u_i)$, where each $u_i$ depends only on the state of agent $i$ and its local neighbors. The associated engineering properties are decentralization, scalability, robustness, and emergent capability: no single point of failure, control laws that persist as agents are added or removed, tolerance to the failure of a few agents, and global behaviors such as area coverage arising from simple local rules (Majee et al., 2024).

A macroscopic line of work seeks low-parameter descriptions of swarm behavior. In the density–performance model, swarm performance is decomposed into a cooperation term $C(N)=a_1N^b$ and an interference term $I(N)=a_2e^{cN}+d$ with $c<0$, yielding a unimodal rise-and-fall curve for PRESERVED_PLACEHOLDER_^^^^2id:(Scherb et al., 2021) OR ti:\2search_query2^^^^. In the urn model of collective decision making, the consensus state PRESERVED_PLACEHOLDER_^^^^2id:(Scherb et al., 2021) OR ti:\2id:(Scherb et al., 2021) OR ti:\2^^^^ evolves under a positive-feedback probability PRESERVED_PLACEHOLDER_^^^^2id:(Scherb et al., 2021) OR ti:\22^^^^, and the expected drift is PRESERVED_PLACEHOLDER_^^^^2id:(Scherb et al., 2021) OR ti:\2^^^^3. These models are presented as candidates for universal properties of swarm performance and collective decisions (Hamann, 2012).

2. Coordination, communication, and formal control

Swarm coordination is usually described through explicit communication topologies and local control laws. In UAV networks, four canonical communication architectures are identified: Star (Centralized), Single-Group Ad-Hoc, Multi-Group Ad-Hoc, and Multi-Layer Ad-Hoc. Routing may be graph-based, using Dijkstra, A*, or Bellman-Ford on PRESERVED_PLACEHOLDER_^^^^2id:(Scherb et al., 2021) OR ti:\2^^^^4, or flooding-based, with controlled gossip as a hybrid compromise between reliability and efficiency. On the control side, the same literature distinguishes leader–follower schemes such as Fixed Global Difference and Double Fixation from fully decentralized consensus laws such as PRESERVED_PLACEHOLDER_^^^^2id:(Scherb et al., 2021) OR ti:\2^^^^5 and second-order protocols for coordinated motion and formation maintenance (Majee et al., 2024).

A more security-oriented formulation appears in SHARKS, a leaderless encirclement protocol for UAVs in smart-city settings. Each agent uses only the target location and the bearing of its nearest neighbor. Motion is the sum of a radial centering term, which maintains the annular constraint PRESERVED_PLACEHOLDER_^^^^2id:(Scherb et al., 2021) OR ti:\2^^^^6, and a dispersion term that rotates the repulsion heading by an angle PRESERVED_PLACEHOLDER_^^^^2id:(Scherb et al., 2021) OR ti:\2^^^^7. The design goal is explicit: organized movement without centralized control, elected leaders, or reliance on GPS, thereby removing a single point of failure (Cooley et al., 2018).

Spacecraft swarms are classified into five classes according to the degree of coordination in motion, sensing, and communication. Class 2 is defined by decentralized motion plus communications coordination: each spacecraft runs its own GNC stack, communication schedules and channels are coordinated, no master node is required, and simple flags such as “in imaging region” versus “out of imaging region” trigger role-specific behavior. This classification is significant because it separates collective operation from full system-level optimization (&&&^^^^2id:(Scherb et al., 2021) OR ti:\2search_query2^^^^&&&).

The formal-methods counterpart is Swarm-STL, which extends STL with predicates over counts of agents rather than agent identities. A Swarm-STL predicate has the form

PRESERVED_PLACEHOLDER_^^^^2id:(Scherb et al., 2021) OR ti:\2^^^^8

so specifications can state that at least PRESERVED_PLACEHOLDER_^^^^2id:(Scherb et al., 2021) OR ti:\2^^^^9 agents satisfy a condition. The key abstraction represents a cooperating group by a centroid plus a bounding ellipsoid, reducing planning from $N$^^^^2search_query2^^^^ dimensions to $N$^^^^2id:(Scherb et al., 2021) OR ti:\2^^^^ where $N$2 is the number of swarms. High-level planning is performed at the swarm level, low-level control preserves containment and inter-agent distances, and the resulting two-stage procedure guarantees satisfaction of the original STL formula while computational time scales with the number of swarms rather than the number of agents (&&&^^^^2id:(Scherb et al., 2021) OR ti:\2id:(Scherb et al., 2021) OR ti:\2^^^^&&&).

3. Programming models, middleware, and experimental ecosystems

One response to the difficulty of engineering emergent behavior is to raise the abstraction level from individual agents to collective fields. MacroSwarm does this through aggregate computing and field calculus. A swarm-behavior block is a pure function mapping sensing fields to actuation-goal fields, written as $N$3. Blocks compose sequentially and in parallel, so flocking can be expressed as the parallel composition of separation, cohesion, and alignment followed by normalization. Because the framework relies on self-stabilizing field-calculus primitives such as rep, nbr, and foldhood, it inherits convergence and resilience properties under asynchrony, message loss, and dynamic topologies (&&&^^^^2id:(Scherb et al., 2021) OR ti:\22^^^^&&&).

ROS2swarm addresses the same problem at the middleware layer. Built on ROS 2 and DDS-based peer-to-peer discovery, it organizes swarm software into movement patterns, voting patterns, and a stand-alone HardwareProtectionLayer. The package exposes ready-to-use primitives for aggregation, dispersion, flocking, random walk, and collective decision making; uses /scan for sensor input and /cmd_vel for motion output; and publishes swarm opinions on a global /opinions topic. It is reported on TurtleBot3 Burger, TurtleBot3 Waffle Pi, and Jackal UGV platforms, with hardware-independent behavior nodes and platform-specific parameter files (&&&^^^^2id:(Scherb et al., 2021) OR ti:\2^^^^3&&&).

At the hardware–software integration level, “Swarm-Enabling Technology for Multi-Robot Systems” retrofits general-purpose robots with a Swarm-Enabling Unit composed of an SBC, a wireless mesh radio, a robot-link interface, and an external power source. Its software library, marabunta, separates body, network, and behavior modules, enabling the same distributed behaviors to be simulated with MockBody and MockNetwork and then deployed on terrestrial robots or marine buoys with minimal code change. The reported platform range includes 45 buoys and 48 robots operating without supporting infrastructure (Chamanbaz et al., 2017).

A larger-scale experimental ecosystem is RISE, the Rapid Integration Swarming Ecosystem. RISE combines a heterogeneous physical testbed of more than 25^^^^2search_query2^^^^ networked agents, a live/virtual/constructive simulation environment, a decentralized MANET, and a gesture-based C2 layer. PyC2 provides tactic authoring and runtime loading, while the “fluid fidelity” mechanism switches agents between low- and high-fidelity simulation based on runtime requirements. In field trials, one operator simultaneously commanded up to ^^^^2id:(Scherb et al., 2021) OR ti:\2^^^^74 agents, exceeding the DARPA ^^^^2id:(Scherb et al., 2021) OR ti:\2^^^^:^^^^2id:(Scherb et al., 2021) OR ti:\2relevance2search_query2^^^^ goal (&&&^^^^2id:(Scherb et al., 2021) OR ti:\2^^^^5&&&).

4. Swarm as a distributed systems architecture

The term also names distributed computing architectures in which coordination, storage, or replication are organized through decentralized swarm-like mechanisms. SwarMS is an ICN-style architecture for infrastructure-less mobile scenarios with three tightly coupled modules: swarm on-boarding, swarm coordination, and a computation executor. Each swarm is represented by an append-only request-log and result-log, each entry carries a hash pointer and a signature, and replicas satisfy a prefix-ordering property. Tasks are replicated across multiple nodes; a node may take a task only when its current count of take-overs or results is below the desired replication factor; and the reliability model is explicit, with success probability $N$4 under $N$5 independent executions (&&&^^^^2search_query2^^^^&&&).

SwarmMesh extends the notion of swarm to distributed data storage in mobile robot collectives without external infrastructure. Each tuple receives a feature-aware key, each robot computes a node score $N$6 from remaining memory and neighborhood size, and key-based routing forwards data toward a storage node capable of holding the tuple. The design goal is to move data away from dangerous periphery locations and toward parts of the swarm better suited to the chosen feature of interest. In simulation, the system maintains near-perfect data retention even in high-load conditions and stores quantities of data that exceed the memory of individual robots (&&&^^^^2id:(Scherb et al., 2021) OR ti:\2^^^^7&&&).

A distinct system named SWARM addresses adaptive load balancing in distributed spatial streaming. It places an adaptivity protocol over tuple-at-a-time engines such as Apache Storm or Twitter Heron, estimates partition cost as

$N$7

and rebalances by moving whole partitions or splitting a heavy partition between the highest-load and lightest-load machines. In experiments on a 4^^^^2search_query2^^^^-node EC2 cluster with ^^^^2id:(Scherb et al., 2021) OR ti:\2^^^^ billion geo-tagged tweets and ^^^^2id:(Scherb et al., 2021) OR ti:\2^^^^ million continuous range subscriptions, SWARM achieved on average a 2^^^^2search_query2search_query2^^^^ % improvement over the best static grid and reduced execution latency on average 4× (Daghistani et al., 2020).

Another SWARM, “Swift WAit-free Replication in disaggregated Memory,” is a replication scheme for shared disaggregated-memory objects. Its two contributions are Safe-Guess, a wait-free single-roundtrip replication protocol, and In-n-Out, a method for conditional atomic update and atomic retrieval of large buffers in one roundtrip. The exported abstraction is a linearizable, wait-free read/write register over memory nodes that provide only one-sided RDMA operations and 64-bit CAS. In SWARM-KV, the measured latency overhead relative to an unreplicated store is marginal, while latency and availability are better than FUSEE; median GET and UPDATE latencies are reported as $N$8 and $N$9, respectively, against $x_i(t)\in\mathbb{R}^n$^^^^2search_query2^^^^ and $x_i(t)\in\mathbb{R}^n$^^^^2id:(Scherb et al., 2021) OR ti:\2^^^^ for RAW (&&&^^^^2id:(Scherb et al., 2021) OR ti:\2^^^^9&&&).

5. Optimization, exploration, and mission design

Swarm research remains closely tied to optimization and path planning. In UAV systems, the surveyed algorithmic repertoire includes PSO, ACO, the Wolf Pack Algorithm, and the Grey Wolf Optimizer, together with PRM, RRT, APF, Dijkstra, A*, D*, and D*-Lite for large-area search. Formation maintenance is expressed as a consensus problem on relative positions, and dynamic reconfiguration is tied to graph connectivity and the Laplacian’s second eigenvalue $x_i(t)\in\mathbb{R}^n$2 (Majee et al., 2024).

SPACE, “Swarm Pheromone Fields for Adaptive Collision-Aware Exploration,” pushes stigmergic coordination to the scale of tens to hundreds of robots. The method maintains a frontier-distance field $x_i(t)\in\mathbb{R}^n$3, an explore pheromone $x_i(t)\in\mathbb{R}^n$4, and a robot-density field $x_i(t)\in\mathbb{R}^n$5 over a shared occupancy grid, and each robot follows a weighted combination of local gradients plus short-range reactive terms. On sixteen HouseExpo home layouts and eight KTH campus floors, with swarms up to 256 robots, SPACE lies on the empirical Pareto frontier: it achieves the lowest inter-robot contact rate at every congested swarm size, from four to seventeen times fewer than a greedy nearest-frontier planner, while keeping coverage time within about two percent of that planner (&&&2^^^^2id:(Scherb et al., 2021) OR ti:\2^^^^&&&).

Adversarial settings produce another optimization problem: swarm-to-swarm engagement. A scalable RL formulation casts the engagement as an MDP whose state is a concatenation of Gaussian-mixture parameters for the controlled and adversarial swarms and whose action specifies the number of groups and Gaussian control-input means and variances for each group. TD3 is used to learn policies that control group steering and density. In the reported curriculum, 8^^^^2search_query2search_query2^^^^ k environment steps converge in about 6 h on a MacBook M^^^^2id:(Scherb et al., 2021) OR ti:\2^^^^, and a single simulation step without training takes about $x_i(t)\in\mathbb{R}^n$6, which the authors present as real-time feasible (Demir et al., 2022).

Mission design for scientific observation appears in the IDEAS software for spacecraft swarms. The Automated Swarm Designer projects feature vertices to flyby visiting points, computes the coverage figure-of-merit

$x_i(t)\in\mathbb{R}^n$7

and runs an evolutionary algorithm to minimize the number of observers subject to a coverage threshold. In the demonstrated Class 2 swarm around asteroid 433 Eros, the design finds a minimum $x_i(t)\in\mathbb{R}^n$8 to achieve $x_i(t)\in\mathbb{R}^n$9 within $\dot x_i=f_i(x_i,u_i)$^^^^2search_query2^^^^, with pointing errors converging to zero within about two minutes after each mode switch and graceful degradation to about $\dot x_i=f_i(x_i,u_i)$^^^^2id:(Scherb et al., 2021) OR ti:\2^^^^ coverage after removal of one observer (&&&^^^^2id:(Scherb et al., 2021) OR ti:\2search_query2^^^^&&&).

6. Applications, analytics, and unresolved issues

Application domains are broad but technically specific. UAV swarms are discussed for disaster management, rescue operations, path planning, and link-budget analysis; spacecraft swarms are motivated by small-body surface observation; SHARKS targets secure autonomous circling of a point of interest in smart-city environments; and “swarm perception” reframes a robot swarm as a support module in a heterogeneous architecture comprising other robots and human operators (Majee et al., 2024, &&&^^^^2id:(Scherb et al., 2021) OR ti:\2search_query2^^^^&&&, Cooley et al., 2018, Kegeleirs et al., 2024).

The support-module perspective directly addresses a recurrent misconception that swarms must be self-sufficient, isolated systems. In that line of work, a swarm of 5 Mercator robots explores an office while a Toyota HSR acts as a client. The HSR issues FIND_PERSON, the swarm shares detected faces and coordinates, and the combined system reduces search time from about $\dot x_i=f_i(x_i,u_i)$2 to about $\dot x_i=f_i(x_i,u_i)$3, achieves $\dot x_i=f_i(x_i,u_i)$4 coverage in 2 minutes, and reaches $\dot x_i=f_i(x_i,u_i)$5 correct face matches versus $\dot x_i=f_i(x_i,u_i)$6 with the HSR’s single camera alone. The proposed concept is “swarm perception”: continuous distributed data collection plus peer-to-peer sharing, exported to external agents through a request–response interface (Kegeleirs et al., 2024).

A parallel line of work asks how a swarm should be characterized by an external observer. “Swarm analytics” organizes individual and collective indicators into an ontological framework of information markers. Markers include speed, ODBA, angular velocity, cohesion, separation, alignment, Transfer Entropy, and Active Information Storage; they are mapped from raw positional data to situations, behavior labels, and context through a recognition-agent pipeline. In shepherding experiments, 42 markers with a decision-tree classifier reach a peak test-set accuracy of $\dot x_i=f_i(x_i,u_i)$7 for agent-type classification and $\dot x_i=f_i(x_i,u_i)$8 for ^^^^2id:(Scherb et al., 2021) OR ti:\2id:(Scherb et al., 2021) OR ti:\2^^^^-way swarm-scenario classification (Hepworth et al., 2022).

Open problems recur across the literature. SwarMS identifies rigorous evaluation under realistic mobility traces, automated tuning of replication parameters, incentive alignment through a “coin” scheme, and security analysis of takeover-and-keepalive dynamics under Byzantine nodes (&&&^^^^2search_query2^^^^&&&). Spacecraft-swarm work points to dynamic clustering, heterogeneous payloads, and realistic RF link budgets, delays, and fault-protection schemes (&&&^^^^2id:(Scherb et al., 2021) OR ti:\2search_query2^^^^&&&). SHARKS leaves moving targets, full 3D encirclement, integrated dynamic obstacle avoidance, and formal convergence proofs under noisy measurements as future challenges (Cooley et al., 2018). Taken together, these directions suggest that swarm research is no longer confined to emergent motion alone; it now includes formal specification, middleware, data management, communication theory, safety guarantees, and system-level integration across heterogeneous platforms.