Examining the Interactions and Dependencies between Active Vision and Reasoning

点击翻译按钮获取中文摘要

The successes of machine learning methods, such as those found in Deep Learning Neural Networks, have transformed computational vision and Artificial Intelligence (AI). Although exciting, in a robotic context, there seems to be uncertainty about how these methods apply. At the 2015 Int. Conf. on Robotics and Automation (ICRA), two presentations emphasize this. D. Rus's (MIT) plenary keynote stressed that "robots can't figure things out": they cannot reason about their circumstances and tasks. G. Hager (Johns Hopkins U), introducing his Workshop on Robotic Vision, highlighted the need for robotic sensing systems to exploit attention and context, key elements for "figuring things out". Perhaps something more is needed. The long-term goal of my research program is to use the language of computation to develop and formalize theories of human vision and visual cognition that have predictive power, thus furthering vision science, and use these to inform the engineering (and not necessarily mimic) of practical robot and machine vision systems. Within this context, the current proposal focuses on how active vision and reasoning strategies may be coupled in order to develop effective active reasoning methods. Our research plan proposes bridging the needs of inference mechanisms to the capabilities of active perception, with a perspective on machine vision that is strongly motivated by a modern understanding of human visual neurobiology. It is our hypothesis that the visual algorithms must be tunable to the input, environment and task of the moment, in a dynamic fashion, in order to enable a system to capture the generality seen in human vision. The goal is to develop computer vision algorithms and demonstrate their performance, testing this basic hypothesis, but being grounded in the robotic perception needs of autonomous companion robots for home, for the elderly, and for the infirm. The importance of advanced robotics research to Canada is easy to demonstrate; NSERC has invested in the Strategic Network for Field Robotics, of which I am a co-PI (PI: G. Dudek). That network addresses both basic research into robotics as well as applications across a wide variety of domains covering terrestrial, airborne, water-based robots for natural resources, health, security and more. Even with this large effort, there are aspects of robotics and robotic perception not covered; this proposal addresses one of those needs. Robots should understand their environment and circumstances, be able to reason about how to perform their tasks, the effects and impacts of their actions, and be able to determine when their actions may not be correct and develop better plans for achieving their tasks. These seem critical not only for achieving robustness in applications in natural resources and manufacturing, but also safety and human-like behaviour for companion robots in the home and for the elderly.