I will present recent advances in visual traffic scene understanding, covering topics such as online action detection, atomic activity recognition, risk object identification, and the corresponding datasets and benchmarks developed for intelligent driving systems. Through our work on algorithms and datasets, we identify a synergy at the intersection of egocentric computer vision, child development, and robotics. This leads us to explore robotic food manipulation, where we develop solutions leveraging active perception, diffusion model-based manipulation policies, and high-fidelity simulation. More recently, we begin investigating how to build computational models that embed emotion into visual experience. In this talk, I will share my findings in these explorations.