- Atlas
- Engineering
- Multi Robot Systems
Multi Robot Systems
Multi-robot systems (MRS) is the branch of robotics concerned with getting more than one machine to work on a task at the same time.
8 frameworks, from 1988 to today · Robotics · Engineering
- Behavior-Based Robotics 1988-2005
- Swarm Intelligence 1992-Present
- Distributed Constraint Optimization 1993-Present
- Decentralized Probabilistic Estimation 1997-Present
- Market-Based Coordination 1997-Present
- Heterogeneous Team Architectures 2000-Present
- Centralized Fleet Management 2003-Present
- Learning-Based Coordination 2010-Present
Overview
Multi-robot systems (MRS) is the branch of robotics concerned with getting more than one machine to work on a task at the same time. The questions it asks are deceptively simple: how do robots share information, divide work, avoid getting in each other's way, and stay coordinated when no single robot can see the whole picture? These questions matter because many tasks—searching a collapsed building, surveying a field, moving goods through a warehouse, flying a light show—are too large, too fast, or too dangerous for one robot. Building a bigger robot is one answer; building many small ones that cooperate is another, and it is the answer this field pursues.
The history of MRS is a sequence of answers to one recurring problem: as the number of robots grows, the amount of coordination any single controller must do grows much faster. Each era of the field found a different way to push that burden down—onto local rules, onto markets, onto probability, onto learning—and each way made larger and more varied teams possible.
From teleoperation to teams: the 1980s
Early multi-robot work grew out of industrial and laboratory settings where several machines shared a workspace. In the 1980s, researchers began asking what happens when robots must cooperate rather than simply coexist. A landmark was the cellular robotic system proposed by Toshio Fukuda and colleagues in the late 1980s: many simple units that could rearrange themselves into different configurations, an early vision of a self-reconfiguring team. Around the same time, Rodney Brooks's subsumption architecture at MIT argued that complex behavior could emerge from simple reactive layers rather than a central world model12. This behavior-based robotics approach was quickly extended to groups: each robot followed local rules, and group-level order was expected to emerge without a leader1. The claim was attractive because it promised scalability and robustness—lose one robot and the group carries on—but it offered few guarantees about what the group would actually do.
Swarm intelligence and the biology-inspired turn
The strongest statement of the emergent view came from swarm intelligence, which took inspiration from ants, bees, and flocks3. In the 1990s, Marco Dorigo's ant colony optimization and the particle swarm optimization of James Kennedy and Russell Eberhart (1995) showed that simple agents following local rules could solve optimization problems collectively. In robotics, the Swarm-bots project, funded by the European Union and running from 2001, built teams of small physical robots that could link together and cooperate on tasks such as moving objects too heavy for one unit. The appeal was scale: thousands of cheap robots, each with minimal sensing, might outperform a few expensive ones. The limitation was control: emergent behavior is hard to certify, and early swarms mostly demonstrated collective motion and aggregation rather than reliable task completion.
Distributed constraint optimization: coordination as a formal problem
A different camp wanted guarantees. If coordination is a problem of assigning tasks and resources among robots subject to constraints, it can be written as a distributed constraint optimization problem (DCOP): each robot holds variables, constraints link variables across robots, and the team must find an assignment that maximizes or minimizes a global objective without a central solver4. Work on distributed constraint satisfaction in the 1990s, followed by DCOP algorithms such as ADOPT (2000) and later DPOP, gave the field formal methods with bounded error and, in some cases, optimality. The price was communication and computation: exact DCOP scales poorly, and much subsequent work went into approximate or anytime variants. DCOP remains the reference point whenever a multi-robot problem needs a correctness argument rather than a demonstration.
Markets and auctions: coordination by bidding
In parallel, a market-based approach treated robots as self-interested bidders. The idea, developed in the late 1990s and 2000s, is that a robot estimates the cost of taking a task, submits a bid, and a simple auction allocates tasks to the lowest bidder. Market-based coordination proved practical for task allocation in teams of tens of robots and connected MRS to multi-agent systems research. Its weakness is that greedy bidding can produce poor global allocations, and much work went into combinatorial auctions and sequential auctions to close that gap. The approach's virtue is modularity: adding a robot means adding a bidder, not rewriting a central plan.
Probabilistic estimation across robots
A quieter but foundational line concerned uncertainty. If each robot's sensors are noisy, the team's shared picture of the world is a probabilistic inference problem. Research on decentralized estimation in the 2000s, including decentralized Kalman filters and later factor-graph methods, asked how robots could fuse their observations without a central node and without double-counting information that had already been shared. The covariance intersection method, introduced by Simon Julier and Jeffrey Uhlmann in 1997, gave a conservative way to combine estimates from sources with unknown correlation. This work underpins modern multi-robot mapping and localization, where teams build a shared map while each robot tracks its own position.
Heterogeneous teams and central fleet management
As robots diversified—ground vehicles, aerial drones, manipulators—researchers confronted heterogeneous team architectures, in which different robot types contribute different capabilities and the coordination problem becomes one of matching roles to platforms. At the same time, industry pushed the opposite direction: centralized fleet management, where a single server assigns tasks to many robots. Warehouse systems such as those deployed by Kiva Systems (founded 2003, acquired by Amazon in 2012) showed that centralized coordination could be enormously effective when the environment is structured and the fleet is large. The tension between centralized control, which is simple to reason about but a single point of failure, and decentralized control, which is robust but harder to guarantee, runs through the field to this day.
Learning-based coordination
From the 2010s, machine learning entered MRS. Reinforcement learning for multi-agent systems, multi-robot policy learning, and imitation learning let teams acquire coordination strategies from data rather than hand-design5. Early successes included learned flocking and formation control, and later work applied deep reinforcement learning to cooperative navigation and manipulation. The attraction is that learned policies can handle complexity that hand-coded rules cannot; the difficulty is generalization, safety, and the credit-assignment problem—when a team succeeds, which robot's action deserves the reward? Learning-based coordination is now a major research front, often combined with the classical methods above rather than replacing them.
Where the field stands
Today MRS draws on all these traditions. Swarm and behavior-based ideas inform large-scale aerial and marine teams; DCOP and market methods remain the tools for task allocation with guarantees; probabilistic estimation is standard for any team that must build a shared model of its environment; heterogeneous architectures and centralized fleet managers dominate commercial deployments; and learning-based methods are increasingly layered on top. The open questions are the old ones in new form: how to certify the behavior of a team whose members cannot all communicate, how to scale coordination without scaling communication, and how to let robots learn to cooperate without learning unsafe habits. The field's history is a record of moving the coordination burden from the engineer to the algorithm—and of discovering, each time, that the burden does not disappear but changes shape.
References
- Behavior-based robotics (History), Wikipedia
- Subsumption architecture, Wikipedia
- Swarm intelligence, Wikipedia
- Distributed constraint optimization, Wikipedia
- Multi-agent reinforcement learning, Wikipedia
The frameworks of Multi Robot Systems
Each framework is a way of working the field has used: when it held, and the question it answers.
- 1988-2005
Behavior-Based Robotics
How can simple local rules produce complex group behavior?
- 1992-Present
Swarm Intelligence
How can simple agents collectively solve optimization problems?
- 1993-Present
Distributed Constraint Optimization
How can a team find an optimal assignment without a central solver?
- 1997-Present
Decentralized Probabilistic Estimation
How can robots fuse noisy observations without double-counting?
- 1997-Present
Market-Based Coordination
How can bidding allocate tasks among self-interested robots?
- 2000-Present
Heterogeneous Team Architectures
How should different robot types be matched to roles?
- 2003-Present
Centralized Fleet Management
How can a single server coordinate a large robot fleet?
- 2010-Present
Learning-Based Coordination
How can teams learn coordination strategies from data?
How the ideas connect
Each arrow runs from an earlier idea to the later one it shaped. The list below says what passed between them.
Swarm Intelligence grew out of Behavior-Based Robotics
Swarm intelligence in robotics grew out of the behavior-based commitment to local rules and emergent group order, replacing hand-designed reactive layers with biological stigmergy and simple agent rules. Early swarm robotics demonstrations of aggregation and collective motion built directly on behavior-based control architectures.
Carried by Marco Dorigo; Swarm-bots project.
Market-Based Coordination reacted against Behavior-Based Robotics
Market-based coordination arose as an alternative to purely reactive, emergent behavior-based control, arguing that task allocation needs explicit economic mechanisms and guarantees rather than local rules. Auction-based systems defined themselves against the unpredictability of emergent approaches.
Market-Based Coordination rival of Distributed Constraint Optimization
DCOP and market-based coordination competed in the late 1990s and 2000s as answers to multi-robot task allocation, one seeking optimality guarantees through distributed constraint solving and the other practical scalability through bidding. Both addressed the same assignment problem with different formal commitments.
Learning-Based Coordination influenced by Swarm Intelligence
Learning-based coordination adopted swarm-inspired tasks such as learned flocking and formation control, using simple local interaction rules as the target behavior for reinforcement learning. Swarm demonstrations provided benchmark problems for learned collective motion.
Learning-Based Coordination influenced by Distributed Constraint Optimization
Learning-based coordination took up DCOP-style task allocation as a benchmark, with learned policies and graph neural networks used to approximate or accelerate distributed constraint optimization solutions. The formal problem formulation carried over into learned coordination research.
Learning-Based Coordination influenced by Decentralized Probabilistic Estimation
Decentralized probabilistic estimation supplied the belief-state representations and uncertainty models that learning-based coordination uses for cooperative navigation and mapping. Learned policies operate over the probabilistic world models developed in decentralized estimation.
Learning-Based Coordination influenced by Market-Based Coordination
Market-based auction mechanisms were incorporated into learning-based coordination as structured allocation layers, with learned bidding policies and reinforcement learning applied to auction-based task assignment. The auction framework provided the allocation substrate that learning methods optimized.
Learning-Based Coordination influenced by Heterogeneous Team Architectures
Heterogeneous team architectures provided the role-assignment and capability-matching problem structure that learning-based coordination later addressed with multi-agent reinforcement learning. Learned policies were trained to allocate roles across different robot types.
Learning-Based Coordination rival of Centralized Fleet Management
Learning-based coordination developed partly as a decentralized alternative to centralized fleet management, aiming to replace the single-server task assignment of warehouse systems with learned distributed policies. The two approaches competed over how to coordinate large robot fleets.
Beyond the field
Where Multi Robot Systems's ideas came from in other fields, and what they led to there: the strongest 12 of 29 links.
Learning-Based Coordination grew out of Multi-agent reinforcement learning in Multiagent Systems
Learning-based coordination of robot teams is largely multi-agent reinforcement learning applied to physical systems: agents learn decentralized policies through shared reward and centralized training. The MARL formulations of joint action learning and credit assignment are the direct source of the later methods.
Learning-Based Coordination grew out of Robot learning in Robotics
Learning-based coordination is a branch of robot learning: it applies the learning architectures, policy representations and training procedures developed for single robots to teams of robots. The later field extends that body of technique to multi-robot settings.
Behavior-Based Robotics grew out of Robotics in Mechatronics
Behavior-based robotics was built directly on the mobile robotics and sensor-based robot engineering of the preceding decades, reusing its platforms and actuators while rejecting deliberative planning.
Carried by Rodney Brooks.
Swarm Intelligence influenced by Behavior-based robotics in Robotics
Swarm robotics and swarm intelligence inherited from behavior-based robotics the reactive, locally interacting agent as the unit of design, applied to groups rather than single robots.
Carried by Rodney Brooks; Marco Dorigo.
Swarm Intelligence influenced by Evolutionary Computation in Artificial Intelligence
Swarm intelligence methods such as ant colony optimization and particle swarm optimization were built as evolutionary and population-based stochastic search techniques, taking their selection-and-iteration scheme from evolutionary computation.
Carried by Marco Dorigo; Russell Eberhart.
Swarm Intelligence influenced by Behavior-based robotics in Multiagent Systems
The multiagent-systems strand of swarm intelligence took from behavior-based robotics the idea of decentralized control through local behavioral rules rather than central planning.
Carried by Rodney Brooks; Marco Dorigo.
Swarm Intelligence influenced by Behavior-based robotics in Artificial Intelligence
Swarm intelligence drew on behavior-based robotics' demonstration that simple reactive agents with local rules can produce robust collective behavior, a model carried into ant-colony and flocking algorithms.
Carried by Rodney Brooks; Marco Dorigo.
Distributed Constraint Optimization influenced by Combinatorial optimization in Operations Research
Distributed constraint optimization applies combinatorial optimization methods, including branch-and-bound and dynamic programming, to problems split among autonomous agents.
Distributed Constraint Optimization influenced by Dynamic Programming in Combinatorial Optimization
Dynamic programming supplied the algorithmic backbone for distributed constraint optimization, notably in dynamic-programming-based message-passing solutions for agent networks.
Carried by Richard Bellman.
Distributed Constraint Optimization grew out of Cooperative Teamwork and Distributed Constraint Reasoning in Multiagent Systems
Distributed constraint optimization grew out of the distributed constraint reasoning and cooperative teamwork programme, formalizing the constraint satisfaction problems of that work as optimization objectives.
Carried by Makoto Yokoo.
Distributed Constraint Optimization grew out of Distributed Artificial Intelligence in Multiagent Systems
Distributed constraint optimization descends from distributed artificial intelligence, inheriting its problem of coordinating autonomous agents that hold private, partial information.
Decentralized Probabilistic Estimation grew out of Probabilistic robotics in Robotics
Decentralized probabilistic estimation extends probabilistic robotics to teams, distributing the Bayesian filtering and mapping machinery of single-robot probabilistic robotics across multiple agents.
Carried by Sebastian Thrun.

Multi Robot Systems is a planet in Nousfarer
Its city has a tower for each of these 8 frameworks. Land, walk in, and climb one concept at a time.
Land on Multi Robot SystemsRelated fields
Fields its ideas are linked to
- Multiagent Systems 11 links
- Robotics 3 links
- Artificial Intelligence 2 links
- Mobile Robotics 2 links
- Robot Learning 2 links
- Robot Motion Planning 2 links
- Embodied Cognition 1 link
- Combinatorial Optimization 1 link
- Operations Research 1 link
- Mechatronics 1 link
- Production Planning And Control 1 link
- Robot Control 1 link
More in Robotics
- Human Robot Interaction
- Mobile Robotics
- Robot Control
- Robot Kinematics Dynamics
- Robot Learning
- Robot Manipulation
- Robot Motion Planning
- Robot Perception And Slam
- Robotics
See where Multi Robot Systems sits on the Map of Human Knowledge