Autonomous robots are increasingly adopted in environments shared with humans, where adherence to social norms is essential for seamless integration. Traditional navigation algorithms, while effective in ensuring efficiency and safety, often neglect the social aspects of human-robot interaction, leading to behaviors that can disrupt human activities or cause discomfort. This thesis presents a novel, unified, learning-based framework for socially aware navigation. Given a map that also encodes people's positions and orientations, a convolutional neural network computes a social cost layer, enabling robots to navigate human environments while respecting social conventions. By integrating social norms into the decision-making process, the proposed approach goes beyond obstacle avoidance to include socially appropriate behavior in dynamic, multi-agent environments.