Faulty Robots

Time-VaryingGraphsandblackholesproblemsdiscusspotentialvariationinthefixed part of the environment: the graph of géométrie environment in which the robots evolve. Problems involving faulty robots introduce the possibility of variations in the mobile part of the environment: the robots themselves. Those variation imply unreliability, which maysignificantlyincrease the amount ofresources requiredfor an algorithm to solve a problem.

The notion ofByzantine faults, that became a central topic in exploration prob­

lems with unreliable robots, was first introduced in [168]. The paper introduces the problem as follows: an army captain must make a decision based on the information he receives from his générais. However, some ofthosegénérais may be liars attempt-ing to make him take the wrong decision. In order for ail loyal générais to reach an agreement, there must be more than two third of ail générais that are loyal. The notion ofByzantine general lias later been applied to the exploration problem in the context of Byzantine faults: a robot may behave in such way as to distribute false information and hinder the accomplishment of a given algorithm. Similarly, asyn-chronous Systems with Byzantineprocesses arestudied in [111], and aweakerversion ofthe problem is discussed in [167].

The study offaulty robots has two principal subdivisions: the crash faults, and the Byzantine faults. In the crash fault model, the robot may be unable to move, communicate or sense its environment properly, but will not willingly hinder the execution of the algorithm for other robots. In the Byzantine scénario, based on [168], [111], and [186], the robot will actively attempt to hinder the other robots by either changing its path, communicating false information or ignoring relevant information.

Most papers studying robots operating in LCM cycles assume that during their Look phase, robots are able to collect perfect datafrom the environment, and during their Move phase, they exactly move at the expected location. [48] drops those assumptions,andconsidersrobotsthatmaybeexpérienceinaccuracyintheirreadings or in their movements. Several impossibility theorems are introduced. It is shown that robots are unable to gather in a finite number ofsteps. An algorithm allowing convergenceis presented, assumingbounded measurement, movement andcalculation errors.

Similarly, [2] considersthe gathering problemfor robotsoperating in LCM cycles, and prove that ail previously known algorithms fail in the presence of either crash faultsorByzantine faults. Moreso, agatheringof3robots, oneofwhichis Byzantine, is impossible in the asynchronous setting. In the synchronous setting, an algorithm solves the gathering problem for k > 3 robots with at most one faulty robot. A general algorithm also solves the problem if there are at least 3/ + 1 good robots, with/ the number offaulty robots.

Gravitational algorithms for robots operating in LCM cycles are studied in [47].

The paper focuses on the asynchronous setting and shows correctness ofthe gravita­

tional algorithm to solve thegathering problem. Analysis ofits convergencerate and resilience to crash faults is also discussed.

Gathering oflabelled synchronous agents in the graph, / ofwhich are byzantine, is studiedin [91]. TwolevelsofByzantine behaviours are studied: astrongByzantine robot can choose its port when moving and convey arbitrary information to other agents, whereas a weak Byzantine robot can do the same, but is unable to lie about its label. For weak Byzantine robots, ifthe size of the graph is known, any number ofgood agents cangather. Ifthe size is unknown, at least /+2 agents must be good agents for the gathering problem to be solvable. This bound is tight. For strong

Byzantine agents, a lower bound of / + 1 good agents is provided, even when the graph is known. Algorithms that ensures gathering are provided and require 2/ + 1 good agents if the size of the graph is known, and 4/+ 2 good agents if the size of the graph is unknown. An open question left in this article is solved in [30|, where it is determined that the minimum number of good agents required to guarantee deterministic gathering ofail good agents, with termination, is /+2.

In [199], a swarm of k labelled robots must solve the gathering problem while being résilient to faults in their sensors and Byzantine agents. For small swarins, the gathering problem is solved and is résilient to any number ofByzantine agents. For larger swarms, an algorithm solving gathering is provided with the assumption that the numberofByzantine agents around a good agent is bounded.

A probabilistic point ofviewof the gathering problem is introduced in [79]. The paper considersagroup ofdisoriented robotsoperating in LCM cycles. Deterministic algorithms are studied under the additional assumption that robots may expérience crash faults or Byzantine faults. A large set ofscheduling strategies are considered and lower bounds are provided.

In [31], a swarm of oblivious robots operating in LCM cycles on the line must solve the convergence problem (that is, be located at a distance that is no more that e apart) despite Byzantine faults. An algorithm allowing gathering for at least 2/+1 goodrobots is providedfor the synchronoussetting, and analgorithm allowing gathering for at least 3/-F 1 good robots is provided for the asynchronous setting, with / the number ofByzantine robots.

Explorationonthelinebyagroupofnrobots,/ amongwhicharefaulty, isstudied in [67]. This paperfocuses on crash-faults. In other words, robots may fail to see the exit, but will not communicate false information. For n > 2/+2, an algorithmwith a compétitive ratio of 1 is provided. For larger /, an algorithm called proportional

schedule algorithm is provided, and is proven to be optimal for n = /+ 1.

A variation ofthepatrolling problemdiscussed laterin thissurvey isdiscussed in [58|. The paperconsidersthepatrollingproblembyagroupof krobots,/ ofwhichare crash-faulty. The environment is aweightedgraph, and the purpose ofthe algorithm is to minimise the visit time by a good robot ofail nodes in the graph. An optimal algorithm is providedfor alinesegment andfor Euleriangraph. For cubicgraph, the problem is shown to be NP-Hard by réduction from the 3-colouring problem.

[201] présents a survey on fault-tolerance and the similar problem offault détec­

tion.

An interesting take on fault tolérance is presented in [130], where k robots prone to faults must explore an infinité sequence ofboxes in order to find a treasure. The authors propose non-coordinating algorithins and discuss contexts in which it may be favourable to implement non-coordinating algorithmsrather than the faster coor-dinating algorithms. Building upon those results, [158] study a variation where the treasure isplaced uniformlyat randomin afinite, large numberofboxes, andpropose algorithms that achieve optimal speed-up.

Resilience to sensorfaults are discussedin [187], where tworobots observing each other and the environment and sharing this information can reduce odometry er-rors and better detect obstacles, thus increasing the quality of the map they create.

Algorithms are introduced and supported by experimental results and simulations.

In [37], the exploration of a network with faulty edges by a single robot is con-sidered. A perfectly compétitive algorithm is provided for ringenvironment, and for networks modelled by Hamiltonian graphs, it is shown that the overhead (the worst-case ratio between the cost ofagiven algorithm and the costofan optimalalgorithm which knows where the faults are located) for a Depth-First Search is at most 10/9 times larger than that ofa perfectly compétitive algorithm.

The notion ofrobotscontrolled by a market economy is used in [209] to solve the exploration problem. Using this approach, the authors allow for dynamic inclusion and departure of robots during the execution of the algorithm. The algorithm they provide is résilient to communication faults.

Other Variations Involving Search by Multiple