Robot embodiments are neither singular nor static. They differ across hardware designs, sensor configurations, and manufacturing variation, and change over time through wear, repair, and redesign. This diversity fragments embodied experience across sensorimotor interfaces, and learning confined to one embodiment blocks the continual improvement that general embodied intelligence requires. This perspective article argues that the response is not to minimize this diversity but to learn from it: to build cross-embodiment robot intelligence, agents that learn from heterogeneous embodied experience and adapt to novel embodiments online. Cross-embodiment learning is then both a means and a goal: a means for accumulating experience across humans, tele-operation devices, simulation, and real robots, and a goal because in telligence must operate across diverse and evolving physical platforms. Realizing it requires understanding what data mixtures, representations, system architectures, and learning algorithms enable transfer and continual learning across embodiments. We synthesize emerging evidence for transfer at semantic, physical, and control levels, and identify open research questions.