Goal-Oriented Task Planning for Composable Robot Control System Using Module-Based Control-as-Inference Framework

Robot control systems built with flexibly combinable sub-task modules can be applied to diverse and complex tasks by reusing or recombining those modules. Previous studies have proposed multiple module designs to achieve a goal task, but they require learning or manual design of dependencies between modules, and in a number of cases, learning the order of execution of modules toward a goal, which is a challenge for a composable robot control system in diverse applications. To solve this problem, we propose goal-oriented task planning based on inter-module similarity. As transitions between modules are possible when the termination of one module is similar to the initiation of another, our method defines the state distributions at the initiation and termination of each module, and automatically estimates the dependencies between modules on the basis of their similarity. We further developed a goal-oriented task planning method by incorporating these modules into the framework of Control as Inference [1], [2]. We evaluated our method in an environment simulating a kitchen task using an arm robot and showed that it performs as well as or better than conventional methods without manual design and learning of dependencies between modules.