Robotic embodiment comprises sensing, kinematics, dynamics, geometry, actuation, and control, through which an agent physically interacts with the world. These properties differ widely across robots and evolve over time. General embodied intelligence therefore requires learning that accumulates across such differences. Existing approaches often engineer correspondences to bridge embodiment gaps, yielding immediate practical gains but imposing assumptions that restrict long‑term transfer scope. A more universal strategy should discover representations that support transfer to a broader range of embodiments as experience grows. We propose treating embodiment diversity as a scaling axis and viewing broad learned priors as a complementary ingredient. At the same time, evaluation protocols are needed to better characterize embodiment gaps and transfer performance. Cross‑embodiment learning not only tackles the engineering challenge of learning from heterogeneous robot experience, but also aligns with a scientific pursuit inspired by nature—physical intelligence that co‑evolves with its embodiments to gain greater agency over behavior and form.
Review