Synthetic data-augmented learning pipelines for cobotic packing work cells

Abstract

Safe human-robot interaction has consistently been one of the main concerns behind industrial robot applications. This is particularly true with the emerging trends in collaborative robotics and their use in quick, relatively inexpensive automation of warehousing and distribution tasks. As such, there is an increasing need for safety features in response to dynamic workspace conditions that were not present in industrial environments in the past. This thesis aims to introduce novel methodologies that allow for the generation of dynamically stable packing pattens, more accurate, comprehensive understanding of 3D scenes from data captured with RGB-D sensors, as well as more energy-efficient and collision free trajectories in collaborative manipulators. The first contribution is based on dynamic stability studies of cutting and packing problems, the next contribution is focused on a new procedure for hand-eye calibration that is not dependent on printed grid patterns. The next addition to the state of the art is related to domain randomization, where approaches towards synthetic data generation and training procedures are proposed. Lastly, a reinforcement learning scheme making use of proximal policy optimization and engineered rewards aiming to reduce inefficient movements in collision avoidant path planning is presented. The mentioned contributions were implemented via a case study in an automated packing operation.