Anytime error recovery by integrating local and global feedback with monitoring task states

Abstract

One of the important points for realizing a plan-based robotic system is to recover from errors while the system executes its plans. In this paper, we discuss a system architecture that is able to perform fault detection and error recovery in any time. The key features of the proposed system are: 1) background refreshing of task-level description, 2) constant updating of geometric representation, 3) anytime monitoring of task states. These features enable it to update a world model description without the task controller and motion executor attending to object perception. Background updating of the world description makes it possible to detect errors as soon as they happen in a portable way. We show two experimental results on the HRP-2 humanoid robot while pouring tea with human interruption. One is an experiment for global error recovery and another is for local error recovery.