Layered Task Allocation in Multi-robot Systems

A layered task allocation method is presented for multi-robot systems in a collaboration and adversarial, dynamic, real-time environment with unreliable communication in this paper. The process of task allocation is divided into three layers: task decomposition layer, task evaluation layer and task selection layer. In task decomposition layer, robots categorize their environments into corresponding modes, and fix subtasks in every mode as experts do, in order to reduce candidate tasks and decrease the complexity of task allocation. Q-Learning based on Adaptive Neuro Fuzzy Inference System (ANFIS) is adopted to compute utilities of candidate tasks in task evaluation layer. This can not only avoid the complicated opponent modeling but also make the learning more efficient. In task selection layer, task with the maximum utility is selected in application, but in learning, task is selected according to randomized Boltzmann exploration tactics in order to get more information for optimization. Simulation experiments implemented on simulated robotic soccer show that this approach improves performances of multi-robot systems greatly.