Machine Learning Concepts for Dual-Arm Robots within Virtual Reality

The collaboration between humans and artificial intelligence (AI) driven robots lay the foundations for new approaches in industrial production. However, intensive research is required to develop machine learning behavior that is not only able to execute shared tasks but also acts following the expectations of the human partner. Rigid setups and restrictive safety measures deny the acquisition of adequate training samples to build general-purpose machine learning solutions for evaluation within experimental studies. Based on established research that trains AI systems within simulated environments, we present a machine learning implementation that enables the training of a dual-arm robot within a virtual reality (VR) application. Building upon preceding research, an activity diagram for a shared task for the machine learning model to learn, was conceptualized. A first approach, using vector distances, led to flawed results, whereas a revised solution based on collision boxes resulted in a stable outcome. While the implementation of the machine learning model is fixed on the activity diagram of the shared task, the presented approach is expandable as a universal platform for evaluating Human-Robot Collaboration (HRC) scenarios in VR. Future iterations of this VR sandbox application can be used to explore optimal workplace arrangements and procedures with autonomous industrial robots in a wide range of possible scenarios.