robotics//fleet//communication constraints

Communication constraints are the limits that real radios and networks place on what the agents of a fleet can tell each other, namely range, bandwidth, packet loss and latency, and they decide which coordination schemes survive outside a simulator. In textbooks a neighbour's information arrives whole, instantly and always; in the field each of the four limits breaks one of those assumptions.

Communication constraints are the limits that real radios and networks place on what the agents of a fleet can tell each other, namely range, bandwidth, packet loss and latency, and they decide which coordination schemes survive outside a simulator. In textbooks a neighbour's information arrives whole, instantly and always; in the field each of the four limits breaks one of those assumptions.

Range makes the communication graph change as agents move. The disk model (a link whenever the distance is below a radius) is convenient; real links are probabilistic and can be one-way.

Bandwidth is shared. Fifty drones broadcasting their state at 10 Hz in 100-byte packets add up to about 400 kbit/s on one channel, which fits on Wi-Fi and does not fit on a long-range telemetry radio carrying tens to hundreds of kbit/s (radio link). If every agent relays everything it hears, traffic grows as N2N^2N2.

Losing 1 to 10% of packets is normal, and the losses come in bursts: a drone passing behind a hangar loses everything for two seconds. Latency runs from milliseconds on Wi-Fi to hundreds on a geostationary satellite link, with peaks far above the mean.

Send state, not increments, and stamp every datum with its time.

A message saying I am at (12.3, 4.1) can be lost harmlessly, because the next one replaces it; a message saying I moved 0.3 m corrupts the receiver's count forever when one goes missing. And a position without a synchronized timestamp cannot be corrected for its age (time synchronization).

Three errors appear that a lone robot never has, and all three travel up into assignment, estimation and control. Stale information: a neighbour moving at 10 m/s whose position arrives 200 ms late is 2 m off, with a perfect GPS. An inconsistent view: two agents with slightly different information go for the same task, or dodge to the same side. And the network partition, in which each group keeps working convinced that the other has vanished.

The rules become configuration in middleware. In DDS, the transport under ROS 2, a state stream suits best-effort delivery that keeps only the latest sample: a lost message is replaced by the next, and retrying it would only deliver stale data. Increments would need reliable delivery, whose retries add exactly the delay a lossy link cannot afford. How these limits shape the split between the centre and the agents is fleet architecture.