Failure-Aware Multi-Robot Coordination for Resilient and Adaptive Target Tracking

2 min read

A team of robots tracks moving targets through an area with unknown “danger zones” that knock out sensing or communication. The framework separates two kinds of failure. Temporary ones (intermittent loss of sensing or links) are handled by inferring where the danger zones are and planning around them, with a tunable acceptable risk. Permanent ones (a robot that has lost sensing or communication for good) are handled by splitting the team: robots that can still talk form a communication group and plan together with partly centralised nonlinear optimisation, while cut-off robots keep working on their own or in smaller groups and stay useful locally. Target estimates come from extended Kalman filters, merged across robots by covariance intersection. A disconnected robot’s behaviour is itself a signal: the others can see where it went silent and infer a communication hazard there. Presented at IROS 2026 in the session “Keeping Robot Teams Resilient Under Adversity”. First posted in August 2025.

Findings

  • Evaluated in simulation over varying numbers of robots and targets, both kinds of danger zone and several failure modes, and on hardware with Crazyflie drones tracking Yahboom ROSMASTER X3 ground vehicles (for example three drones and three targets).
  • With four robots, four targets and one unknown sensing danger zone: under a risky setting, both this method and the baseline suffer serial sensor failures, and tracking error climbs fast after about 120 steps. Under regular and conservative settings the robots keep a finite estimation error as they follow the targets. The authors say the point is the range of behaviours, not the tracking error.
  • The abstract reports consistent robust performance but gives no headline number.

Limits

  • Failures are loss of capability, not wrong data. A robot that keeps talking but reports bad estimates is outside the model.
  • Danger zones are static and parameters are hand-tuned. Mobile danger zones, evading targets and partial degradation are future work (authors).

Relevance

The split between “can still coordinate” and “on its own” is a practical form of graceful degradation, close to resilient task allocation and resilient coverage. Its use of a teammate’s silence as evidence about the environment is worth keeping in mind for failure-recovery in teams. Because covariance intersection trusts every estimate it merges, it would need something like resilient-consensus or link-level trust once faulty data, not just silence, is on the table. See robust-robotics.