Proactive-reactive detection and mitigation of intermittent faults in robot swarms

3 min read

Oğuz and colleagues go after intermittent faults: errors that appear, vanish and come back, such as GPS that degrades sporadically in clutter. These are hard to catch in ad hoc swarm networks, so the paper builds on self-organised hierarchical networks with a leader (the SoNS approach, with hierarchical Henneberg construction). Proactively, robots run an adaptive biased minimum consensus protocol to keep a backup path to the leader up to date as the topology changes. Reactively, each robot compares the positional information arriving by the primary and backup paths with a one-shot likelihood ratio test. When the test fires, traffic is rerouted through the backup path until the fault clears.

Findings

  • The consensus protocol is proven to converge to a minimum-cost backup path on piecewise-constant dynamic graphs. With 5 to 9 of 20 robots (25 % to 45 %) removed at random, it always found a backup path, usually within one or two hops of the primary.
  • Detection in 20-robot formations with channel bit error 0.02: with one faulty parent, accuracy rose from about 22 % to about 90 % as fault probability increased; with two faulty parents, from about 34 % at fault probability 0.01 to over 90 % by 0.30. Mean accuracy across fault rates was 61.3 % (one parent) and 75.6 % (two parents), with mean false positive rates of 17.9 % and 9.2 %.
  • Against a centralised Bayes-optimal likelihood-ratio benchmark, the method lost 10 % to 30 % accuracy at very low fault rates or heavy noise and came within a few points at moderate rates.
  • In a 200-robot formation at background fault probability 0.1, the method held formation almost perfectly while the unprotected swarm broke down within 25 s. A 20 s burst at 0.35 cost about half the robots, and more than 30 were still in formation at 100 s.

Limits

  • Only offset faults. Stuck-at and spike faults are future work (authors).
  • Assumes some fault-free path exists in the multiplex network and that the leader is fault-free (authors).
  • Needs a persistent hierarchical network, which fully ad hoc swarms do not have.
  • Simulation only. Faults are random. A relay that corrupts both paths consistently, or colluding relays, are not considered (my reading).

Relevance

A rare paper that pairs detection with a concrete recovery (rerouting) and measures what it buys. Formation control under uncertainty is also the topic of the buoy swarm workshop paper. See failure recovery and resilient consensus; the gradual-fault counterpart is okeeffe-2025-antibody-swarm-faults.