Toward Visual Communication Methods for Underwater Human-Robot Interaction

Abstract

Trained divers take on the complex and often dangerous underwater environment to perform essential tasks. These tasks include inspection and repair of underwater infrastructure and monitoring the health of water systems through tasks such as observations of coral reefs and tracking of invasive species. Autonomous Underwater Vehicles (AUVs) able to assist with these tasks have become more widely deployed as their capabilities improve, however, when deployed as solo agents they lack the intuition and ability to adapt to unexpected situations as a human diver would. The objective of collaboration between a diver and an AUV brings together the ability of an AUV to perform tasks that are dangerous to the human diver, while maintaining the ability of the diver to monitor the situation and update task information as necessary. For this collaboration to be successful meaningful communication is essential, especially when the goal of the collaboration is to complete a task. This dissertation presents our work towards improving diver-AUV collaboration, focusing on utilizing visual perception onboard the AUV. In the following chapters, we discuss two novel communication algorithms that allow divers to communicate information about the location of an object required by an AUV to perform a task. These methods have been designed to take into account challenges such as limitations of on-board computation as well as challenges inherent to working in the underwater domain, such as non-traditional human body poses and limitations of traditional, terrestrial, computer vision. Evaluations of these methods are performed onboard AUVs. We then incorporate these algorithms into a communication system which allows a diver to assign the AUV a task based on the object detected. This system also provides feedback from the AUV to the diver about the task which will be performed, forming a closed loop communication system between diver and AUV. Validation of this system was performed fully onboard an AUV in the Caribbean Sea. In addition, as AUV visual perception can be hampered by the visual degradation of the underwater environment, we therefore present an investigation into a task-based method to improve AUV vision. We also discuss our contributions to the design and creation of the research platforms necessary for this research to move forward.