Simulation of a palletizing robot in NVIDIA Omniverse Isaac Sim with a computer vision system

This thesis presents the design and implementation of a robotic palletizing simulation framework developed in NVIDIA Omniverse Isaac Sim. The work addresses the gap between high-level pallet layout generation and the low-level execution constraints of an industrial manipulator operating in a cluttered environment. In the proposed workflow, pallet configurations are provided by an external planning module, while the simulated robotic system is responsible for perception, grasping, optional package reorientation, collision-aware transfer, and stable placement on the pallet. The developed environment acts as a digital twin of an industrial palletizing cell. It includes a KUKA manipulator model, a surface gripper, a simulated overhead RGB camera, JSON-driven order generation, and a physics-based execution pipeline. The perception module estimates the pose of incoming packages through classical image processing, while motion execution is governed by an RMPflow-based controller with different operating modes for free-space transfer and contact-sensitive phases. Two task-specific modules complement the controller: a package reorientation routine for packages that must be rotated before placement, and a placement strategy that combines pallet-layout post-processing with directional insertion logic to reduce lateral impacts during dense stacking. The main contribution of the thesis is the integration of these components into a coherent manipulation pipeline that maps industrial order data into a complete perception-to-placement loop. The thesis does not claim a full experimental benchmark on real hardware; rather, it demonstrates that a structured simulation framework can be used to study the interaction between packing logic, perception uncertainty, motion generation, and contact-aware placement before deployment in a physical cell.