Resilient Task Allocation in Heterogeneous Multi-Robot Systems

2 min read

Mayya and colleagues reassign a team of aerial and ground robots to tasks when weather, terrain or an attack degrades specific capabilities. Each robot compares the predicted and measured change in a scalar task-value function. A central “mission evaluation” block maps these discrepancies onto capability degradation scores (ground mobility, aerial mobility, sensing resolution) using a species-to-capability matrix. An event-triggered mixed-integer quadratic program then picks, for each task, which combination of capabilities to use, assigns robots, excludes robots degraded past a threshold, and relaxes the requirements of lower-priority tasks as little as possible when the team cannot meet them all.

Findings

  • In a CoppeliaSim scenario with three tasks (track target 1, track target 2, cover an area), a low-friction patch immobilises the ground robot on task 2. The planner flags ground mobility, switches task 2 to an aerial-only configuration, sends two aerial robots, and returns the stuck robot to its start.
  • At 27 s a “fog” event degrades sensing on task 2. With no spare robots left, the planner sacrifices the lower-priority coverage task and moves one aerial robot to tracking.
  • Over 20 randomised trials with 32 aerial and 8 ground robots, the capability margins stayed at or near zero with reallocation and went negative without it.
  • With every optimisation parameter perturbed (standard deviation 40 % of its mean) over 20 trials, results stayed close to nominal.
  • The MIQP solved in under one second for large teams on a laptop.

Limits

  • Allocation is centralised and needs a link to the central computer (authors). A distributed version is future work.
  • Degradation is attributed by simple averaging within a task, and the authors suggest fault detection and diagnosis methods would do better.
  • The model assumes deviations in task value come only from disturbances in the dynamics. Attacks enter only as unmodelled disturbances, so a robot that misreports its own state would be misread as degraded capability (my reading).
  • Simulation only.

Relevance

One of the few papers in this cluster where detection feeds a recovery action other than “stop”: the team reconfigures and degrades gracefully. See failure recovery in multi-robot systems, and robust robotics more generally.