There is solid empirical evidence that both geometric and semantic visual representations are crucial for executing actions in embodied intelligent systems. Thus, solving visual perception is a pressing need, as the large-scale deployment of such intelligent systems can greatly benefit humanity, e.g. through autonomous cars or robotic home assistants. However, the Physics-agnostic data-driven approach taken by the vast majority of recent works on visual perception exhibits fundamental limitations in generalization, often yielding physically infeasible and inaccurate outputs. By contrast, in this talk I will advocate for a hybrid, data-driven yet Physics-informed approach to embodied visual intelligence. This approach embraces the well-proven representational power of generic parametric neural network mappings, while at the same time it organically integrates fundamental principles of optics and radiometry in the learning. In this way, we can inject physically valid inductive biases to the learned models and grant them broader capability of generalization. The merit of our Physics-based regime will be empirically supported by showing it in action in a range of exemplary works of ours on visual perception, which address the representative problems of metric 3D depth estimation, inverse rendering, and multi-sensor fusion for semantic segmentation and neural reconstruction.