Effective robotic manipulation hinges on the ability to encode compact, structured representations of the environment, particularly the spatial relationships among objects. Prior work on spatial reasoning has largely relied on supervised learning with human-annotated or heuristic labels. In this paper, we introduce an unsupervised object-centric representation learning framework that infers spatial relations directly from raw visual inputs without relational labels. Our method leverages a disentangled variational autoencoding architecture combined with adversarial training to enforce the suppression of object-specific features while preserving inter-object spatial information in joint embeddings. We train the model on synthetic scenes containing object pairs with known spatial configurations and validate the learned representations on downstream classification tasks. Empirical results demonstrate that our approach yields superior generalization to unseen object shapes and colors, and significantly outperforms baselines trained directly on pixels or using standard autoencoding methods for spatial reasoning in manipulation contexts.