Team-Aware Multirobot Strategy for Cooperative Path Clearing

In this paper, we present a simulated version of a potential demining problem in which robotic minesweepers clear a battle area of anti-tank mines to enable troops to breach the field. The demining problem is modeled as a distributed optimization problem in which the robots strive to minimize an abstract cost function. Robots were simulated using the Java based simulator, TeamBots (www.teambots.org). The demining robots are responsible for clearing a direct vertical path between the top and bottom edges of the minefield. At each time step, robots must select one of three possible actions: move, scan, or defuse. Each action has a time cost associated with it that represents the relative difficulty of the task for the robot; deactivating mines is not explicitly modeled in the simulation as an actuation task. This problem has characteristics of the consume and graze tasks described in (Balch & Arkin 1995); robots explore unknown areas until they isolate a path that has few mines before converging to “consume” the mines. By varying the costs associated with each action, we can model the characteristics of different robotic deminers. Previous work suggests that a homogeneous strategy should perform well in this domain, given its similarity to other foraging-type tasks. However we show that robots using a simple homogeneous approach have difficulty completing the task due to high inter-robot interference. We suggest that the proper way to approach the problem of interference is to make the the robots more “team-aware”; each robot plans its strategy based on assumptions about what its teammates are doing and current sensor data. This approach works well and substantially outperforms the simple homogeneous approach in which each robot individually optimizes. Although this type of global optimization can become intractable for large numbers of robots, we present an efficient algorithm that only considers local interactions. For the “team-aware” optimization strategy, each robot models the decision-making process of its nearest peers. Since robots outside this neighborhood are less likely to interfere the robot’s goal, their impact on the choice of next action is not considered. For each teammate within this neighborhood the robot determines whether the teammate is also heading for its chosen column. This calculation can be performed in O(K) time for K neighbors and creates a