Detecting and diagnosing faults in autonomous robot swarms with an artificial antibody population model

2 min read

O’Keeffe targets gradual wear (motors losing efficiency, sensors losing range) instead of the sudden failures most swarm fault work injects into a few robots. In the Artificial Antibody Population Dynamics (AAPD) model, each robot keeps populations of artificial antibodies tied to patterns of behaviour (“paratopes”). Populations grow when a pattern matches what a neighbour is doing and decay otherwise, following Farmer and colleagues’ antibody dynamics, with a self-tolerance term. The dominant population says which robot is faulty and which part (left motor, right motor, sensor) is degrading. A flagged robot goes back to base for repair while it can still get there. The model is distributed and has supervised and unsupervised versions.

Findings

  • With every robot degrading stochastically, a foraging swarm using AAPD worked on average at 70 % to 97 % of its fault-free performance and avoided robots failing in the field in most cases tested.
  • With complete motor or sensor failure in one robot of 10, the zeroth-order model detected every case with no false positives (Fβ = 1). The author compares this informally with published detectors whose sensor-failure Fβ scores were 0.4 to 0.7 in a different setup with 20 robots.
  • Because every robot degrades, there is no guaranteed healthy majority, which breaks the usual data-driven assumption that “normal” is whatever most robots do.

Limits

  • Simulated TurtleBot3 robots in ROS 2 and Gazebo, at most 20 robots, a foraging task, and batteries lasting five simulated minutes at full draw.
  • Comparisons with other detectors are informal, not head-to-head (author).
  • False positives rise in some scenarios when the repertoire of faulty paratopes grows (author).
  • The model does not separate internal faults, external hazards such as rough terrain, and Byzantine behaviour. The author lists this as future work.

Relevance

A rare closed loop of detection, diagnosis and recovery for swarms, aimed at long-term autonomy, where wear builds up over weeks. It belongs with failure recovery and pairs with oguz-2025-intermittent-faults on intermittent faults.