Optimal Navigation System for a Mobile Robot to Execute Dynamical Multiple Social Tasks

Best Master Thesis of NTUEE in 2019
Shao-Hung Chan [pdf]

Abstract

In recent years, researches related to social and companion robots have gradually increased, showing its importance in the field of daily healthcare and human companion. Those robots also demonstrate potential applications especially in the society where elderly people growing year by year. In order for robots to provide assistance toward family members and elders in a household environment, the prerequisite capabilities are to perform robust localization, navigation, and sensing ability. In addition to that, the robots should also be capable of perceiving the environment and human beings based on the visual and audio sensor data. In other words, robots should know how to estimate human status and understand his/her verbal commands so as to complete social and service tasks in the area of intelligent human robot interaction. More practically, a dynamic, any-time decision making system is necessary for social and companion robots to generate adequate task and motion planning (TAMP) over a long period of time. On the other hand, with the purpose of making robots widely deployed in the future, efficient calculation under limited computation resource should be taken into consideration while designing the overall system.

In this thesis, inspired from the Dynamic TAMP framework, we propose a novel task-oriented navigation system for robots to achieve social interaction tasks with the help of perceptions. To organize these social tasks, we propose an instruction structure consisting decaying reward with regard to priorities and time. Moreover, we model the indoor scenario into a graph structure to allocate instructions, and propose a task planning algorithm that considers not only the priorities among multiple tasks but also time efficiency through optimizing the accumulative reward. As for the perceptions that help assign priorities of instructions, we propose a sub-system for human localization, identification, and framewise hierarchical activity recognition in the visual aspect. As for verbal perception, we design a sub-system to understand human words as well as sentiments. Note that under the limited computational speed and resource, the system aims to simultaneously perform perception and decision making using both deep learning modules and heuristic algorithms. With the help of our system, the social robot is able to not only meet human requirements but also interact with people in a multiple-human environment efficiently, achieving sophisticated human robot interaction (HRI).