control//multi-agent control

Multi-agent control is the branch of control that designs the behaviour of several agents, each running its own loop of sensing, deciding and acting, whose loops are coupled through what they tell each other, so that the group achieves what no single agent could: covering a field, holding a formation, agreeing on a meeting point or on an estimate. A **multi-agent system** in this sense is a set of physical agents coupled by a shared objective (cover the area together), by space (do not collide), by resources (one charger, one radio channel) or by information. It is a different object from the language-model agents of multi-agent system in ML, which share the name.

Multi-agent control is the branch of control that designs the behaviour of several agents, each running its own loop of sensing, deciding and acting, whose loops are coupled through what they tell each other, so that the group achieves what no single agent could: covering a field, holding a formation, agreeing on a meeting point or on an estimate. A multi-agent system in this sense is a set of physical agents coupled by a shared objective (cover the area together), by space (do not collide), by resources (one charger, one radio channel) or by information. It is a different object from the language-model agents of multi-agent system in ML, which share the name.

The family exists because the joint problem explodes: ten robots with five actions each have 510≈105^{10}\approx 10510≈10 million joint actions per step, the curse of dimensionality raised to the number of agents, and local rules on a graph sidestep that count. The object all the members share is the communication graph (who hears whom) and its graph Laplacian. Most members are one rule seen under different conditions: each agent moves its value toward its neighbours' values.

The speed and the fragility of every member are read in one spectrum.

The second Laplacian eigenvalue λ2\lambda_2λ2​ (spectral gap) sets how fast the group agrees and is zero when the group is split; the largest one bounds the step size and the delay the group tolerates.

The consensus protocol is the base rule: every agent averages with its neighbours, and the group converges to the mean with nobody computing it. It is worth its cost when no node may be indispensable or the topology keeps changing; with a reliable leader on a good link, broadcasting a reference is simpler and faster.

Formation control runs consensus on offsets, so the group holds a shape while agreeing where to put it. Leader-follower adds one agent that listens to nobody and pulls the others to its value, which is also how a single stuck or malicious agent hijacks the group; resilient consensus is the defence.

Flocking uses three local rules (separation, alignment, cohesion) and lets collective motion emerge, the base behaviour of swarm robotics.

Truck platooning is the application closest to the road: vehicles share speed and braking over a radio link and control their spacing cooperatively, so they can drive closer together. It is at the stage of research and pilots.

On a real drone the multi-agent layer runs slowly (1 to 10 Hz) and hands references to the vehicle's own control at about 50 Hz, which works because the timescales are separated (time-scale separation); collision avoidance is a separate layer with priority. Which decisions belong to a centre and which to the agents is the subject of fleet and fleet architecture, and agreement in the computing sense is distributed consensus.