Cognitive comprehension framework for human-centered situation learning and adaptation in robotics, A

Abstract

Human-centered environment, which is defined by robots, human, and environmental conditions, is a key part of robot task executions. Accurate understanding of human-centered environment is the precondition of successful robot executions in real-world situations. However, in practical situations, there are a lot of environment uncertainties, such as task execution dynamics, tool/human user varieties, temporal/spatial limitations and scenario unstructured characteristics. Robot task execution performances have been largely undermined when robot task execution goes from controlled lab environments to uncontrolled practical environments. To improve robot execution performances in practical human-centered environments, in this dissertation, a three-layer cognitive framework is designed to support comprehensive robot understandings for dealing with environment uncertainties, making robot to “think” like a human, instead of merely to “act” like a human. With the cognitive comprehension framework, mainly three contributions have been made: 1). by abstracting low-level executions and real-world observations of human behaviors, robot behaviors, and environment conditions, high-level cognitive understanding is generated from a human perspective, endowing robots with abstract understanding of human-centered situations, 2). by flexibly decomposing a high-level abstract goal into low-level execution details, robots are able to flexibly make plans and revise plans according to human requirements and environment condition limitations, and 3). the three-layer cognitive framework is updated and evolved as more robot commonsense knowledge is learned. In this dissertation research, this framework is cooperated with efficient robot knowledge learning methods, such as web-mining supported knowledge collection and learning from demonstrations, supporting adaptive robot executions with different domain knowledge.