Robel: synthesizing and controlling complex robust robot behaviors

Abstract

Abstract. We briefly present the Robel supervision system 3 which learns from experience robust ways to perform high level tasks. Each possible way to perform a task is modeled as a Hierarchical Tasks Network whose primitives are sensory-motor functions. The relationship between supervision states and the appropriate modality is learned through experience as a Markov Decision Process (MDP). This MDP is independent of the environment and characterizes the robot abilities for the task. Presentation Robust robot navigation is a complex task which involves many sensory-motor (sm) functions such as localization, path planning, terrain modeling, motion generation adapted to obstacles, and so on. Since no single method or sensor has a universal coverage, each sm function has its specific weak and strong points. The approach presented here improves the global robustness of complex tasks execution