Cross-embodiment

Cross-embodiment refers to training or transferring robot policies across multiple physical platforms — different degrees of freedom, different actuator types, different sensor suites. The strict version: one policy that runs unmodified on platforms A, B, and C. Looser versions: policies trained on multi-platform datasets and then fine-tuned per platform; policies that share representations but use platform-specific action decoders.

The distinction matters because cross-embodiment is the central scaling claim of foundation-model-for-robotics work. Physical Intelligence's π0 and π0.5 papers, Google DeepMind's RT-X and Open X-Embodiment dataset, and Skild AI's brain models all claim cross-embodiment generalization. The strict-vs-loose split mirrors the zero-shot generalization split: a model that requires per-platform fine-tuning is making a meaningfully weaker claim than one that doesn't. Cross-embodiment without methodology details defaults to the loose version.