Embodied AI systems learn from continuous, first-person visual streams — long-form video that provides naturalistic spatiotemporal grounding, multimodality, and object motion, but also constant distribution shift. Iconic single-subject image datasets don't prepare models for this regime. We develop methods for embodied representation learning, world-model prediction and planning, unsupervised continual video learning from egocentric streams, and incremental recognition in open-world environments. Our vision is efficient, adaptable, on-device visual learning that grounds perception in real-world action — with applications in embodied planning, robotics, and visual assistance.