Predictive Fault Tolerance for Autonomous Robot Swarms

TL;DR

This work proposes a predictive approach to fault tolerance, based on the principle of preemptive maintenance, in which potential faults are autonomously detected and resolved before they manifest as failures, and is shown to improve swarm performance and prevent robot failure in the cases tested.

Abstract

Active fault tolerance is essential for robot swarms to retain long-term autonomy. Previous work on swarm fault tolerance focuses on reacting to electro-mechanical faults that are spontaneously injected into robot sensors and actuators. Resolving faults once they have manifested as failures is an inefficient approach, and there are some safety-critical scenarios in which any kind of robot failure is unacceptable. We propose a predictive approach to fault tolerance, based on the principle of preemptive maintenance, in which potential faults are autonomously detected and resolved before they manifest as failures. Our approach is shown to improve swarm performance and prevent robot failure in the cases tested.

Figures (4)

• Figure 1: A: A screenshot of the simulated tunnelling swarm with highlighted maintenance and recharging zone, digging corridor, excavation zone, communication range of an individual robot (approximately to scale), and arrows indicating the path taken during the excavation algorithm. B: A high level state machine providing an overview of our proposed excavation algorithm. Avoiding collisions, maintaining an unbroken communication chain, and returning to the recharging/maintenance area when needed will take priority over the basic excavation algorithm. C: Relationships between degradation functions and degradation severity coefficients listed in \ref{['table:symbols']}.
• Figure 2: Row i shows the total number of blocks excavated by the 5-robot swarm in 15 minutes of simulated time, when up to 5 robots suffer from A: sensor degradation, B: excavation degradation, or C: motor degradation without our PFDDR system; or D: sensor degradation, E: excavation degradation, or F: motor degradation with our PFDDR system. Row ii shows the total power consumed by the 5-robot swarm in 15 minutes of simulated time as a percentage of a single robot's battery capacity, where columns A-F indicate the same fault categories, with and without our PFDDR system.
• Figure 3: A high level state machine outlining our proposed PFDDR system. Each constant is selected with respect to the equations given in \ref{['table:symbols']} and the rates of power consumption observed during experiments for $dc_E < 0.3$ and $dc_{l,r} < 0.3$.
• Figure 4: A shows a comparison in the total number of blocks removed by the swarm when the swarm operates in ideal conditions, with faults in combination and without the implementation of our predictive PFDDR system, and with faults in combination with the implementation of our PFDDR system. B shows a comparison in the total power consumed by the swarm as a percentage of the total battery capacity of a single robot when the swarm operates in ideal conditions, with faults in combination and without the implementation of our PFDDR system, and with faults in combination with the implementation of our PFDDR system. C shows the value of degradation severity coefficients at the time a fault was detected.