Chapters

4 Multi-Robot Systems: Taxonomy, Estimation, and Coordination

robotics
coordination
physical-world

4.1 Overview: A Team Adds Information Paths, Not Just Robots

Robot teams, swarms, and networks can divide work or cover places far apart. They can cover more ground, split jobs, and keep working if one robot fails. At times they can do more together than alone. The lecture gives search and rescue, surveillance or monitoring, and product pickup or delivery as examples.

More robots also add coupling. An estimate made on one robot can influence another robot’s estimate, a delayed message can arrive after the state it describes, and a coordinator can become a single point of failure. Multi-robot design therefore begins with taxonomy, not with a fleet-size target.

4.1.1 Three Independent Taxonomy Axes

Architecture

Centralized systems route estimation or control through one unit. Decentralized systems distribute those decisions across robots.

Communication

Explicit communication sends state that teammates cannot directly observe. Implicit communication occurs when robots infer information from observable motion or environment changes.

Heterogeneity

Homogeneous teams share capabilities. Heterogeneous teams deliberately combine complementary sensors, mobility, payload, or compute.

These axes are independent. A heterogeneous team can still be centrally controlled. A decentralized team can use explicit radio messages. A homogeneous swarm can coordinate partly through observable motion.

4.1.2 Topology Shapes Coordination Risk

Fully Connected

Every robot exchanges information with every other robot. Information paths are short, but message count and dependency management grow quickly.

Star

All robots communicate through a central node. Coordination is simple to inspect, but the center and its links become critical dependencies.

Random Mesh

Each robot communicates with a changing neighborhood. The topology tolerates some failures but makes connectivity and convergence time-dependent.

The goal of decentralization is to approach the performance of an ideal centralized system without inheriting its single point of failure. The costs are communication delay and overhead, asynchronous inputs, and solutions that may be temporarily sub-optimal. Some algorithms degrade gracefully under whatever communication is available, but that property must be demonstrated for the chosen algorithm and topology; it is not automatic.

4.2 Practitioner: Share Relative Observations Without Counting Them Twice

Distributed estimation asks each robot to estimate a local or global variable from incomplete evidence. Kalman filters and particle filters can be distributed by fusing relative observations of teammates with communicated position beliefs.

Collaborative localization provides the concrete workflow:

1. Observe

Robot 1 measures relative range and bearing to Robot 2.

2. Exchange

Robot 2 communicates its position estimate and uncertainty.

3. Transform

Robot 1 maps the relative observation into its local frame.

4. Fuse

The filter updates weights or state while preserving dependency history.

In a particle-filter version, the detected robot reweights its particles using the detecting robot’s belief and the relative measurement, then resamples from the new weights. A Kalman-filter version can perform the same conceptual fusion when its model assumptions fit.

Rumor-propagation warning: a relative observation is not independent merely because it arrived in a new message. If Robot 1's estimate influenced Robot 2, and Robot 2 later sends that estimate back as if it were fresh independent evidence, the same information is counted again. Confidence grows without new evidence. Correct distributed fusion tracks dependency or information history instead of treating every received estimate as independent.

Log at least the observation timestamp, source robot, frame transform, source-estimate version, covariance or particle-belief summary, and dependency identifier. Without lineage, a filter cannot distinguish a new measurement from recycled information.

Particle filter narrated tour

Play silently with captions on provisional storyboard timing; narration audio is pending. Production pacing is not robot or sensor timing. Choose Reduced motion for a complete still at each beat.

Play the tour to open the workbench. Pause, replay a beat, or read the full transcript below. Team observations and information history cannot be shown by this estimator; those beats keep the existing frame and provide captions only.

Full transcript — Particle filter (8 beats)

B1 — Rescue robot uncertainty
You are supervising connected robots on a search and rescue mission. They exchange position estimates to help locate one another. Your concern is confidence growing without fresh information. How can a robot narrow its possible positions while keeping track of where its information came from?

Model caption: Your role: supervise connected rescue robots. Position estimates carry uncertainty; counting the same information again can inflate confidence.

B2 — Particles and team observations
A particle filter represents uncertainty with many possible positions, called particles. Each particle has a weight describing its importance. In the chapter's robot team, a relative distance and direction measurement combines with a teammate's position estimate. The animation explores the particle update within that larger coordination problem.

Model caption: Each particle is a possible position; its weight is its importance. The chapter combines relative distance and direction with a teammate's position estimate. The workbench illustrates the estimator portion.

B3 — Equal weights and prediction
Start with the robot localization example and one hundred twenty particles. Initially, all have equal importance. Divide one by one hundred twenty: each weight is one over one hundred twenty. Prediction moves the possible positions through a motion model and spreads them with motion uncertainty. Their weights stay unchanged during prediction.

Model caption: 120 particles start with weight 1/120 each. Prediction moves and spreads positions; weights remain unchanged.

B4 — Measurement and weighting
Next comes a sensor reading. A likelihood is a score for how well a possible position fits that reading. Nearer positions receive larger scores. The scores become shares of a total of one, called normalized weights. The robot's possible positions now have unequal importance.

Model caption: A likelihood scores fit to the reading. Nearer particles score higher. Normalized weights share a total of one; position importance changes.

B5 — Starting ESS arithmetic
Effective sample size, or ESS, measures how concentrated the weights are. Work through the equal starting weights: square the fraction one over one hundred twenty, add all one hundred twenty squares, then divide one by that sum. The result is one hundred twenty. At the fifty percent threshold, multiply one hundred twenty by one half: sixty. The starting cloud needs no resampling.

Model caption: Starting weights: 1/120 each. Square: 1/14400. Sum 120 squares: 1/120. Divide 1 by that sum: ESS = 120. Threshold: 120 × 0.5 = 60. The initial cloud does not trigger resampling.

B6 — Resampling and equal weights
After weighting, an effective sample size less than sixty triggers resampling. Systematic resampling selects copies according to the weights, favoring stronger possibilities. Small random shifts spread the copies slightly. With one hundred twenty particles again, divide one by one hundred twenty to restore each equal weight.

Model caption: Automatic trigger: weighted ESS strictly less than 60. Systematic resampling favors higher weights; roughening adds small random shifts. The 120 resulting weights are each 1/120. A forced demonstration, if used, does not indicate that the threshold was crossed.

B7 — Recycled information
The misconception is that a new message always brings independent information. Suppose the first robot's estimate influences the second, which sends it back as fresh information. The same observation gets counted twice, increasing confidence without anything new. For your rescue team, tracking the information's history is part of combining position estimates.

Model caption: A new message can contain an old observation. First robot → second robot → first robot can count the same information twice. Shared information history belongs to team fusion; the workbench does not model it.

B8 — Share information history
Refine possible positions with measurements, and keep their information history when sharing the result.

Model caption: Refine possible positions with measurements, and keep their information history when sharing the result.

4.2.1 Choose the Coordination Problem Before the Algorithm

Multi-robot coordination is several problem families, not one:

Problem familyGoalMechanisms named in the source
Coordinated motionmove as a formation or flocklocal potential fields, graph/network control, leader-follower
Allocationdistribute roles or resourcesmarket-based and threshold-based algorithms
Coveragespread robots over a spatial areaLloyd’s algorithm and gradient-based coverage control
Agreementmake local states convergeconsensus over neighbor interactions

Do not choose consensus for a problem that is actually allocation, or use flocking rules where a coverage guarantee is required. First name the variable to coordinate: pose, velocity, role, resource, territory, estimate, or clock.

4.3 Under the Hood: Neighbor Rules Become Team-Level Behavior

Graph control abstracts the communication network as nodes (robots) and edges (communication links). Robot ii has a neighbor set Ni\mathcal{N}_i determined by the current topology or a proximity rule. The controller then uses local states from that set rather than assuming global access.

4.3.1 Flocking: Three Local Steering Rules

Reynolds’ boids model (1987) produces flock-level motion from three local rules:

Separation

Steer away from nearby neighbors to prevent crowding or collision.

Alignment

Steer toward the average heading or velocity of neighbors.

Cohesion

Steer toward the local center of neighboring positions so the group remains together.

Each robot reacts only to neighbors within a region of influence defined by distance and angle. Team behavior changes when that neighborhood graph disconnects, when sensing is asymmetric, or when delay makes neighbor velocity stale. The rule names do not guarantee collision freedom; gains, limits, update timing, and geometry still matter.

4.3.2 Consensus: State the Local Update Contract

The lecture closes with the general consensus update associated with Olfati-Saber, Fax, and Murray (2007):

xi[t+1]=f(xi[t],{xj[t]∣j∈Ni}).x_i[t+1]=f\left(x_i[t],\{x_j[t]\mid j\in\mathcal{N}_i\}\right).

The function ff is often an averaging rule. Robot ii updates from its current value and the values available from its neighbors. The intended outcome is decentralized agreement: all robots converge to the same value through local interactions. Motion coordination, cooperative estimation, and synchronization all fit this form once xix_i is chosen appropriately.

For a simple weighted average:

xi[t+1]=wiixi[t]+∑j∈Niwijxj[t],x_i[t+1]=w_{ii}x_i[t]+\sum_{j\in\mathcal{N}_i}w_{ij}x_j[t],

with non-negative weights summing to one. The update stays within the local convex hull of received values. Convergence still depends on the graph and update assumptions: the network must retain enough connectivity, weights must be well formed, and delays, packet loss, or adversarial values cannot be ignored.

4.3.3 Worked Agreement Trace

Three robots start with scalar values 00, 66, and 1212. On a fully connected graph, let every robot replace its value with the average of all three values:

xˉ=0+6+123=6.\bar{x}=\frac{0+6+12}{3}=6.

All three reach 66 in one update because every node sees the whole graph and uses identical weights.

On a line graph, each end robot sees only the center as a neighbor. If every robot synchronously averages its own value with all its neighbors’ values, the first update is ((0+6)/2, (0+6+12)/3, (6+12)/2)=(3,6,9)((0+6)/2,\ (0+6+12)/3,\ (6+12)/2)=(3,6,9). The next update is (4.5,6,7.5)(4.5,6,7.5): agreement takes multiple updates.

On a disconnected graph, separate components can converge internally to different values; local agreement is not global agreement.

The source lecture states exponential-rate convergence as the consensus outcome. Treat that as a property of a qualifying algorithm under its stated graph and timing assumptions, not as a promise attached to any function named consensus.

4.4 Summary

View a robot team as robots linked by paths for sharing information. First ask who makes decisions, how robots share information, and how their abilities differ. Trace each shared observation: record its time, frame, uncertainty, and history. Then ask what the team must do. Motion, task allocation, coverage, flocking, and consensus each deal with a different goal. Each also rests on its own timing and link assumptions.

An update equation alone cannot make a team agree. Name the paths that carry data and the delays and failures the method can bear. State how received values are protected. Say how each robot responds if links break or data becomes stale.

  • Multi-robot architecture, communication mode, and heterogeneity are independent taxonomy axes.
  • Decentralization removes one central failure point but introduces asynchronous information, delay, overhead, and temporary sub-optimality.
  • Collaborative localization fuses relative range or bearing with communicated beliefs, while preserving measurement dependency history.
  • Coordinated motion, task allocation, coverage, flocking, and consensus solve different coordination variables.
  • Consensus turns local-neighbor updates into agreement only when topology, weighting, timing, and integrity assumptions support convergence.

4.5 Key Takeaway

A robot team acts on shared information.

State who can observe whom and which evidence has been fused. Track changes in the neighbor graph. Name the assumption that lets local updates produce the desired team behavior.

4.6 See Also

Follow the links from local truth toward team behavior. Kinematics defines what each robot can actually do, while classical control explains how each platform arbitrates its own behaviors. Kalman filtering supplies the predict-update and uncertainty discipline needed before estimates are exchanged. The UAV chapters then show those contracts in a mobile network where trajectories, radio links, neighbor graphs, and mission roles can all change together.

Hi, I’m Motor Max. What do you need?

Find help

Let’s find what you need

Find a lesson or return to your module guide.

Give feedback

Help improve this chapter

Tell us what helped or what needs fixing.

Tips

Try a small next step

A few ways to get more from this chapter.

Ask AI

Continue the conversation

Take the conversation further with your preferred assistant.

AI help is available in supported modules when enabled in Settings.

No live person answers on independent study. Feedback does not promise a personal reply.

For welfare or safety concerns, follow your school’s procedures outside IoTClass.