Approaches for Autonomous Multi-Vehicle Underwater Inspections

Abstract

Underwater robotic inspections are an ongoing challenge for large naval platforms. Although many approaches to specific aspects of robotic inspection tasks have been previously explored, the limitations of specific approaches leave large gaps in underwater inspection capabilities. By developing and implementing a cooperative multi-robot system for underwater inspections, we aim to bridge these gaps in existing methods. Cooperative underwater robotic systems face the unique challenges of highly dynamic environmental conditions, extremely limited communications and sensor ranges, expensive and easily damaged hardware, and limited simulation options. This dissertation focuses on bringing together variations on existing methods for task allocation, task discovery, SLAM, and motion planning, modifying them to fit the unique underwater challenges, and integrating them into a system that takes advantages of the similarities and strengths of different methods to build a cooperative robotic system for underwater inspection and maintenance. To begin the testing and integration of various task allocation, SLAM, and motion planning, these methods must first be tested on simulations and datasets. By leveraging aspects of several tools, we are able to develop a basic simulation that allows testing of coordination methods while allowing a controlled environment with the basic tools needed to evaluate the performance of a decentralized task allocation method. By modifying the d-CBBA method to account for task additions and communication limitations, we enable our simulated robots to distribute dynamic sets of tasks. Without prior knowledge of the geometry of the objects being inspected, a method to generate appropriate inspection locations during exploration is needed to determine which tasks should be passed on to be allocated between the robots. With the extreme limitations on communications, there is no way to communicate environment maps between robots in real-time. Individual robots can generate their own maps through onboard SLAM processes, but the only environmental information they receive from other robots in the system is the location of tasks passed between robots. To approach the task discovery problem, we develop Map-TIDAL, which uses the individual robot maps to generate candidate tasks, then compares them to the set of existing tasks made available to the robots through the task allocation to determine the best tasks to add. By integrating task allocation, SLAM, and task discovery, the robots are able to coordinate their efforts with significantly reduced communication bandwidth. By adding motion planning methods into the Map-TIDAL framework, the multi-robot system will be able to perform inspections without risking collisions that could damage the robots or the vessels being inspected. Analysis is performed on several available planners, to determine which planners would be most applicable to the underwater inspection domain.