Intelligent service robots are increasingly being deployed in households, hospitals, warehouses, and other environments. These robots are capable of performing complex tasks, largely relying on their vision systems for object detection. As these robots operate and interact in 3D space, data obtained with 3D sensors is crucial for object detection, as it offers a detailed representation of the environment in the form of point clouds. These point clouds allow for precise extraction of information about the position, size, and geometry of objects within the 3D environment. Examples of such sensors include LiDARs and RGB-D cameras. State-of-the-art 3D object detection methods are predominantly deep learning-based and, as such, require large amounts of annotated data to perform effectively. However, acquiring such data can be both expensive and time-consuming. As a result, synthetic data generation presents a promising alternative, but current object detectors trained on synthetic data still underperform compared to those trained on real data. This suggests that current synthetic data generation methods do not yet achieve sufficient realism. This thesis aims to address this issue by exploring the field of realistic synthetic 3D data generation for indoor environments. The primary goal is to analyze the impact of five key factors of realism: presence of background objects, camera noise, positioning of objects, context of scenes, and object sizes. To investigate these factors, a modular method for generating realistic synthetic single-view point clouds was developed. This method allows for the generation of large, customizable datasets with varying levels of realism, specifically controlling for the aforementioned factors. These datasets are used to train state-of-the-art object detectors, and the impact of each realism factor is evaluated based on the detection performance. Furthermore, the experiments show that the performance of an object detector can be improved by pretraining it on a baseline synthetic dataset and fine-tuning it on real data. Notably, the model trained on geometric data only using this approach outperforms the same object detector trained solely on real data, which uses both geometric and color data. In addition to detecting objects, a service robot needs the ability to interact with objects in their environment. One particular challenge arises with openable objects such as cabinets, nightstands, and closets. While traditional openable objects often feature simple handles that can be grasped to open them, modern design increasingly incorporates handleless doors and drawers. This presents a significant challenge for service robots, as most existing methods for the robotic manipulation of articulated objects focus on objects with handles, and there is limited research on handling objects without them. This thesis addresses the first crucial step in managing such objects: the classification of the opening mechanism. Three categories are proposed for this classification: objects with regular handles, objects that can be grasped by their surface to be opened, and objects with a push latch mechanism. Given that the latter two categories often appear visually similar and can be difficult even for humans to distinguish based solely on appearance, this work explores methods that utilize both images of the objects and images of a human demonstrating the approach to opening them. Experiments conducted using a new dataset indicate that combining images without demonstration and images with demonstration significantly improves the performance of CNN classifiers compared to using either type of image alone. Additionally, experiments conducted with a modern object detector show that such classifiers can be applied to automatically detected regions providing evidence that the proposed methods could be used in real-world environments as part of fully autonomous systems.