1) A single robot has limited sensors (typically a few meters range) and can communicate only with immediate neighbors. 2) Each unit runs the same set of simple control rules (e.g. keep distance X, move toward neighbor average, avoid collisions). 3) There is no central planner — no robot has a global map or identifiers of other units. 4) Local interactions give rise to collective behavior (formation, aggregation, gradient following, collective transport). 5) Algorithms are often bio-inspired (ant pheromone trails → ant colony optimization, herd motion → flocking/Boids) or based on potential-field control. 6) Multi-agent reinforcement learning (MARL) is increasingly used to learn swarm rules.
Tasks too large, dangerous, or geographically distributed for a single robot — disaster exploration, long pipeline inspection, forest protection, large-scale agriculture — require many units operating in parallel. Centralized control of n robots does not scale (communication bandwidth, single point of failure, latency). Swarm Robotics solves this through decentralization: each robot decides locally, and global behavior emerges from scale.
A small, autonomous unit with limited sensors and modest compute. Runs the same rule set as the rest of the swarm. Deliberately designed to be cheaper and simpler than a classical mobile robot — unit complexity is traded for swarm scale.
Official
An information exchange mechanism limited to immediate neighbors (radio range, infrared, light signals, virtual pheromones). No global broadcast, no central broker. Local communication is the key constraint that enforces decentralization.
Official
An identical set of simple rules applied by every swarm member — from the classical Boids rules (separation, alignment, cohesion) to policies learned via MARL. The rule set must guarantee emergent global behavior despite the absence of global state.
Behavior observed at the swarm level that cannot be directly read off the rules of a single unit. Classical examples: flock formation, collective transport, aggregation at a light source, cluster segregation.
Rules that on paper should yield behavior A in practice yield B. Collective behavior is hard to predict analytically from local rules — it requires extensive simulation and empirical validation.
Naive broadcasts in large swarms (>100 units) quickly saturate radio bandwidth and cause packet collisions. A classical beginner mistake in swarm projects.
A swarm should be robust to losing individual units (one of the paradigm's main strengths), but in practice ill-thought rules can lead to cascading failures (a cluster of dead robots blocking information flow).
Craig Reynolds publishes Flocks, Herds, and Schools: A Distributed Behavioral Model — a mathematical bird flock model based on three rules (separation, alignment, cohesion). The foundation of swarm algorithms.
Gerardo Beni and Jing Wang introduce the term Swarm Intelligence in the context of cellular robotic systems — the wider family to which the later Swarm Robotics belongs.
Erol Şahin formalizes the Swarm Robotics field in his classical article, establishing the criteria that distinguish a swarm from an arbitrary multi-robot system.
Michael Rubenstein and team at Harvard demonstrate a self-organizing swarm of 1,024 Kilobot microrobots. The first practical demonstration of swarms at the thousand-unit scale (Science, 2014).
DARPA launches OFFensive Swarm-Enabled Tactics (OFFSET) — a military program developing tactics for swarms of 250+ autonomous drones. A signal that swarms are moving from the lab to operations.
The international AURA Foresight consortium (UK + Australia) reaches the final of the XPRIZE Wildfire Autonomous Response with a drone-swarm system for wildfire detection and suppression. Practical civilian application across a 1,000 km² area.
Each unit executes its rules independently, asynchronously and in parallel. Active compute paths depend on local sensor state and neighbors (input-dependent). Globally the system is sparse in the informational sense — each robot sees only a small slice of the world.
Full execution parallelism — n robots execute their rules in parallel without a coordinator. Training swarm policies (e.g. MARL) is also massively parallel in simulation, typically on GPU clusters with tens of thousands of parallel environments.
By definition, swarm robots are designed cheap, light and with modest compute — microcontrollers (Kilobot: Atmel), Raspberry Pi (Crazyflie), Jetson Nano. Swarm scale compensates for unit simplicity.
Training swarm policies via multi-agent reinforcement learning (MARL) requires GPU clusters in simulation — analogous to whole-body action foundation models.
