Accurate 6-DoF pose estimation for Unmanned Aerial Vehicles (UAVs) is crucial for autonomy but challenged by issues like GPS-denial and the difficulty of acquiring large-scale labeled real-world data. This paper investigates the feasibility of training Deep Neural Networks (DNNs) exclusively on synthetic RGB-D data for drone pose estimation, targeting applications such as indoor navigation and pose tracking (distance, azimuth, elevation). We propose SlicingCNN3D, a custom architecture utilizing sparse convolutions on voxelized point clouds derived from RGB-D input, and compare it against a Point Cloud Transformer (PCT) baseline. Models were trained solely on synthetic data generated in Blender, with and without noise augmentation, and evaluated on real-world data captured using a ZED 2i camera observing a Parrot ANAFI drone. Our results demonstrate successful sim-to-real transfer, with noise augmentation proving beneficial for robustness. The proposed SlicingCNN3D, if trained with noise, achieved promising real-world performance, yielding an overall positional error ( \(e_{XYZ}\) ) of approximately 11.7 cm and notably outperforming PCT in Yaw estimation (12.1 \(^{\circ }\) vs. 30.7 \(^{\circ }\) error), a critical parameter for flight direction. Our findings validate the synthetic-only training approach as a viable pathway of developing capable DNNs for UAV pose estimation from RGB-D sensors.