This paper presents a learning-based direct regression approach for the 6D pose estimation from monocular RGB images, targeting real-time robotic applications. A dataset is generated using a robot-camera setup with a two-stage process involving ArUco marker-based labeling and marker-free image acquisition. A convolutional neural network is proposed to regress the 3D translation and 4D rotation, represented as a unit quaternion, from single RGB images. The model is trained and evaluated on the collected dataset, which includes over 10,000 image-pose pairs of a liquid hydrogen fuel connector relevant to a maritime robot-assisted e-fueling process use case. The network is deployed on a soft programmable logic controller (PLC), and inference times are measured and compared between real-time single-core and non-deterministic multi-core processing. Pose estimation accuracy reaches up to 98.9% for a 4°, 4 cm threshold, and inference times below 750 ms are achieved. The results demonstrate that the learning-based direct regression approach needs further hardware acceleration in order to meet real-time requirements while maintaining a sufficient level of pose estimation accuracy for robotic applications.