What is cross-embodiment robotics?
Cross-embodiment robotics is the study and engineering of methods that allow knowledge, policies, representations or skills learned on one physical platform to transfer to another. The target may be a different robot arm, humanoid, mobile manipulator, dexterous hand or simulated embodiment.
The central challenge is simple: the task can stay the same while the body changes. A grasp learned by a five-finger hand cannot be copied directly to a parallel gripper. A locomotion policy for one humanoid cannot assume identical joint limits, link lengths or actuator dynamics on another.
Why embodiment creates a transfer problem
Robot observations and actions are tied to hardware. Cameras have different viewpoints, proprioceptive states have different dimensions, and action interfaces may expose joint positions, torques, Cartesian motion or high-level commands. Geometry and dynamics further change the consequences of an action.
Successful cross-embodiment systems therefore seek abstractions that preserve what is reusable while adapting what is body-specific.
Common approaches
- Shared representations: encode tasks, objects, language or motion in spaces that are less dependent on a single morphology.
- Retargeting: translate demonstrations or trajectories from a source body into the kinematics of a target body.
- Multi-embodiment training: train on data from many robot types so a model learns common structure.
- Embodiment conditioning: explicitly provide the model with robot morphology, sensors or action-space information.
- Simulation and augmentation: vary robot bodies and dynamics to improve generalization before deployment.
Why it matters for Physical AI
General-purpose physical intelligence cannot scale efficiently if every new robot requires a completely separate model and dataset. Cross-embodiment learning offers a path toward reusable robot intelligence: shared knowledge at the model level, with an adaptation layer for each physical platform.
This is especially relevant to humanoids, where different manufacturers expose substantially different kinematics, hands, sensors and control stacks.
Related topics
Explore robot embodiment, cross-embodiment learning, robot foundation models and VLA models.