Spatial relationships among objects are fundamental to the structure of human environments and play a critical role in enabling effective human-robot collaboration (HRC). In domains such as homes, kitchens, and offices, social robots must be capable of understanding and reasoning about different spatial relations to interpret human intentions and execute tasks in a meaningful way. Learning these relations from perception and language presents unique challenges, including ambiguity in human instructions, variability in object configurations, and limited training data for rare spatial patterns. A Bayesian approach offers a principled way to model uncertainty in both perception and relational reasoning, allowing robots to make more robust predictions in unfamiliar scenarios. By capturing uncertainty over learned representations, Bayesian Neural Networks (BNNs) can reliably generalize to novel spatial arrangements and support safer, more interpretable robot behaviors during collaborative activities. In this work, we propose a novel deep learning model based on BNNs for Spatial Learning (BNN-SL), a framework that can recognize objects and learn the spatial relations among them. Experimental results demonstrate that BNN-SL effectively captures uncertainty and relational structures, empowering robots to execute collaborative tasks that require spatial generalization and compositional reasoning.