Automatically detecting surface defects from images is an essential capability in manufacturing applications. Traditional image processing techniques are useful in solving a specific class of problems. However, these techniques do not handle noise, variations in lighting conditions, and backgrounds with complex textures. In recent times, deep learning has been widely explored for use in automation of defect detection. This survey article presents three different ways of classifying various efforts in literature for surface defect detection using deep learning techniques. These three ways are based on defect detection context, learning techniques, and defect localization and classification method respectively. This article also identifies future research directions based on the trends in the deep learning area.

Many missions require unmanned surface vehicles to operate in complex environments over large distances. Marine environments are dynamic in nature and change over time as a result of tides, weather, and environmental restrictions. As a result, the available free traversal space for a marine vehicle in the given map can change over time. This requires the marine vehicle to dynamically generate and update the traversal map to represent the free space. This paper presents techniques for speeding up A* search on the nodes of visibility graphs. We use a quadtree representation of the marine environment to enable the algorithm to efficiently compute the nodes of the visibility graphs. We have developed an admissible heuristic that considers the large islands while estimating cost-to-go, and provide better estimates than a Euclidean distance-based heuristic. Due to the large size of the marine environment, the branching factor of the search tree can become large and reduce the efficiency of the path planner. We have developed an approach that focuses the search by only considering the child nodes lying in the local region around each node. Our experiments demonstrate that by focusing the search, we can reduce the computational time without significantly sacrificing the optimality of the computed path. We also present an extension to the algorithm for handling dynamic environments with time-varying free spaces. We incorporated the depth charts developed by the National Oceanic and Atmospheric Administration (NOAA), Silver Spring, MD, USA, along with the tide predictions to estimate the time-varying free space that can be used by the planner.

Conventional material extrusion additive manufacturing (AM) processes require the user to make a trade-off between surface quality and build time of the part. A large bead filament deposition can speed up the build process; however, it leads to surfaces with high roughness due to the stair-stepping effect. The surface quality can be improved by using a small bead filament deposition, which increases the build time of the part. We present a new approach incorporating hybrid multi-resolution layers in material extrusion additive manufacturing to provide excellent surface quality without increasing the build time. Our slicing algorithm generates planar layers with large filaments to fill the interior regions in less time. The generated exterior layers are conformal and use small filaments to reduce the stair-stepping effect and improve surface quality. We also present a path planning algorithm to build parts with a single manipulator using a multi-nozzle extrusion tool. The path planning algorithm generates a smooth material deposition path by avoiding collision between the tool and the already built layers. It reduces the collision checks and performs collision detection in a computationally efficient manner. We build five parts to validate our approach and illustrate the benefits of multi-resolution AM.

Mobile manipulators can be used for machine tending and material handling tasks in small volume manufacturing applications. These applications usually have semi-structured work environment. The use of a fully autonomous mobile manipulator for such applications can be risky, as an inaccurate model of the workspace may result in damage to expensive equipment. On the other hand, the use of a fully teleoperated mobile manipulator may require a significant amount of operator time. In this paper, a semi-autonomous mobile manipulator is developed for safely and efficiently carrying out machine tending tasks under human supervision. The robot is capable of generating motion plans from the high-level task description and presenting simulation results to the human for approval. The human operator can authorize the robot to execute the automatically generated plan or provide additional input to the planner to refine the plan. If the level of uncertainty in some parts of the workspace model is high, then the human can decide to perform teleoperation to safely execute the task. Our preliminary user trials show that non-expert operators can quickly learn to use the system and perform machine tending tasks.