Where to Look? Probabilistic Reasoning Methods to Locate Workpieces in Unstructured Manufacturing Environments

• Main Author: Rajaraman, Mabaran
• Format: Dissertation
• Language: English
• Published: ProQuest Dissertations & Theses 01-01-2020
• Subjects: Mechanical engineering; Robotics
• ISBN: 9798678190239

Knowledge of workpiece location is a critical piece of information in any manufacturing industry. Starting from storage, a workpiece's location knowledge is necessary for quick pulls before it gets fixtured and setup (staged) for further tasks to be performed. After a workpiece has been staged, depending on the task tolerance, pose information of the workpiece is necessary to successfully execute the task. However, when uncertainty creeps into the workpiece's location or pose, delays are introduced in the execution of actions involving the workpiece, resulting in reduced productivity for the industry. Progress in sensor and localization technology has allowed mass manufacturing industries to build product-specific structured environments that maintain precise workpiece location information at all times. However, such structured environments are not enforceable in a variety of industries due to economic and practical limitations stemming from the product type and environmental setup. These industries operate in unstructured environments that cannot maintain precise workpiece location at all times. In this work we consider the problem of locating a workpiece both before it has been staged as an inventory tracking problem, and after it has been staged, as a workpiece localization problem in unstructured environments. For the case of workpiece localization, we first consider the task of welding for products with short lifecycles. This task is executed manually in current industrial practices due to uncertainty in the pose of a workpiece resulting from manual general-purpose fixturing. In this thesis we propose a setup that investigates the value of robot-guided manual fixturing coupled with automated workspace inspection by a 6-DOF robot to localize and execute the welds on the workpieces. From experiments, we show that our proposed approach can improve throughput of such welding tasks by ~85%. Next, we consider the task of workpiece assembly in an unstructured environment with a mobile robot. In this case, we show that staggered localization of the workpiece, with varied tolerances, needs to be considered when approaching it to perform tasks. While current vision-based approaches in literature present a single best-fit solution for workpiece pose, they do not consider inherent appearance ambiguities that might exist for the workpiece. This limits the ability of a robot to plan actions and improve its belief when ambiguity in the appearance of a workpiece exists due to similarity in appearance across viewpoints. In this thesis we present a method to learn the similarity in the appearance of a workpiece across views from rendered images. We show that this model can be used to present a distribution of likely poses for any given observation allowing the robot to further take actions to resolve its position with respect to the workpiece. While localization of workpieces once they have been staged for a task is an active research topic in the field of robotics and industrial automation, locating workpieces as inventory before staging has garnered little research interest. In this thesis, we investigate inventory tracking in unstructured environments like shipyards and construction industries. In these industries, poor pull times for workpieces due to disjoint actions by multiple stakeholders is a well documented problem. Current suggested solutions focus on methods to restructure the environment to adopt mature sensing technology like Radio Frequency Identification (RFID) to track workpieces. However, due to several technical limitations and challenges in adopting RFID based tracking, we propose a probabilistic reasoning based system that can hypothesize possible locations of a missing workpiece to assist in the otherwise exhaustive search. Our proposed system collects critical information through viewpoints of various stakeholders, identifies critical events, and reasons the effect of such events to estimate the location of missing workpieces. In our experiments, we show that our system can present likely locations for a missing workpiece and reduce the search effort by ~60% in terms of distance traveled and ~80% in terms of visited workpieces when compared against exhaustive searches. We further show that the proposed probabilistic reasoning based system can present information that expands its scope as a virtual shop-floor system. We present methods to extract actionable information that can further assist in searches. Besides providing assistance to search, we show that our model can be used to estimate the effort to search for any workpiece in the shop-floor. We also make a case to use this information to better plan tasks, and take actions as preventive maintenance to limit or check the uncertainty associated with workpiece locations.