“Dynamic analysis and assistance in the world of work assembly”
Leif Goldhahn, Christina Pietschmann and Katharina Müller-Eppendorfer
Abstractt— Today’s working environment in the field of assembly is characterized by high product variance and dynamically changing work tasks and contents. Artificial intelligence (AI) offers many opportunities and can shape the interaction between humans and technical equipment. Workers can be supported and assisted in ergonomic self-reflection, perception, analysis tasks and in the evaluation tasks of their activities. Therefore, the focus is also on health protection and the preservation of the workers’ working ability within their tendentially longer working life. This article shows potential areas of application for data-based assistance systems in the field of manual assembly. It especially considers the dynamic evaluation of workstation and assembly situations. By combining rule-based data acquisition and data evaluation with a machine learning approach, an evaluation and generation of communication data for logistical or assembly activities can take place in near real-time. Thus, healthy, and productive work is promoted.
Keywords: AI, digital assistance, ergonomic evaluation, ergonomics, work assembly
“Attention guiding techniques in augmented reality when interacting with real and virtual objects”
Svetlana Wähnert and Yuchen Zhang
Abstractt— Visual search in augmented reality (AR) can become more efficient by using attention guiding techniques (AGTs).
Studies have shown that the different AGTs have not only general but also context-specific advantages. How attention can best be directed to a search object depends, among other things, on the features of that object. A key feature of objects in AR is whether they are real or virtual. In this work, we compared the search performance between the 3D visual arrow and the attention funnel depending on the search object (real vs. virtual). We also investigated whether an object-specific combination of these two AGTS would lead to performance gains. In a controlled experiment, participants searched for either real or virtual objects and were supported with different AGTs: (1) 3D visual arrow only, (2) attention funnel only, or (3) object-specific combination of these two. The hypothesis that attention funnel has advantages especially concerning virtual objects could not be confirmed. Instead, attention funnel showed object-independent superiority. Furthermore, the object-specific combination did not bring any performance gains compared to a general use of the two AGTs.
Keywords: Attention guiding techniques, Visual arrow, Attention funnel, Visual search task
“Ubiquitous Semantics for Open World Robotics”
Sascha Griffiths, Alexander Perzylo and Florian Röhrbein
Abstractt— This paper presents a concept of a hybrid embodied AI for robotic systems that draws from experiences in formal knowledge representation, brain-derived embodied technologies and sharing of knowledge between robots. Semantic modeling of robot tasks, interaction objects, and related environments may provide the required level of context understanding to automate data-driven learning pipelines and to handle open world complexities. An increased maturity of these technologies will enable the translation to real applications in future hybrid societies.
Keywords: Knowledge Representation, Hybrid Systems, Embodied AI
“MLOps for Building a Human-in-the-Loop System for Continuous Learning in Mobile Robotic Vision”
Raja Judeh, Jude Ng, Christian von Reventlow and Florian Röhrbein
Abstractt— Mobile robots usually require continuous monitoring during operation for safety reasons. Additionally, teleoperation plays an important role in this regard, where a human can interrupt the robot and take control in case of failures. In the absence of a standard method, the vast amount of data collected by the robot may grow intractably during deployment. This paper aims at introducing the idea of adopting the MLOps framework to exploit teleoperation and continuous monitoring of mobile robots to improve robotic vision. By adopting this workflow, a human-in-the-loop system can be created for continuous learning where teleoperators can strategically and efficiently improve robotic perception models throughout all the steps of the
machine learning life cycle.
Keywords: MLOps, Human-in-the-Loop, Teleoperation, Perception, Mobile Robots