ActionPool : a novel dynamic task scheduling method for service robots

Abstract

Service robots require the seamless utlisation of several technical disciplines. Most of the required technologies are sufficiently advanced to provide feasible solutions to be used in the designing of service robots. For instance, mechanical engineering, control theory, electronics and electrical engineering aspects of the design have all matured well. On the other hand, it is the perception and artificial intelligence that provide the means for modelling the environment and the knowledge which are lagging behind. The latter two disciples in their current state, greatly limit the complexity of the tasks which can be performed by service robots. In this thesis, an ActionPool method for representing task knowledge and executing multiple tasks simultaneously with service robots is presented. The method is based on a concept in which the actions that are ready for execution are placed into a pool and from those most suitable for the situation are selected one by one. The number of actions in a pool and the number of tasks are limited only by the available computational resources. The actions can belong to different tasks, and thus the action pool allows the robot's indivisible resource to be dynamically dealt out for various tasks requiring the resources. In the ActionPool method, the functional parts of the service robot are divided into resources and an action pool is assigned to each one of them. This way, numerous tasks can be executed simultaneously. The ActionPool method allows a natural way of dynamically adding and removing tasks to and from the robot's active execution. The action selection method can direct the perception processes to observe the relevant parts of the environment. The ActionPool method has been implemented on two different service robot platforms to verify the generic nature of the method. Several tasks have been executed successfully to validate the claims about the qualities of the method. Compared to previous approaches, this work provides a fresh execution- and contingency-centric vantage point to the well studied robot control problem.