Cross-Embodiment Transfer

Cross-embodiment transfer studies how a robot can reuse skills or data from a different body, including transfers from human demonstrations to robots. Common approaches align end-effector motion, kinematics, contact patterns, or latent actions, and use shared policies, diffusion models, vision-language-action models, and morphology-conditioned world models. Work tests transfer in settings such as tabletop manipulation, humanoid whole-body control, locomotion, and navigation, using held-out robot embodiments in simulation and on physical robots. A central challenge is making transfer work when both appearance and dynamics change: zero-shot methods can fail on unseen bodies, while even small amounts of target-robot data can help, and benchmarks are still clarifying what counts as genuine transfer.