What is the embodiment gap?
The robot embodiment gap is the difference between the physical and computational assumptions of a learned system and those of the robot on which it must operate. The gap appears whenever a policy, dataset or model moves across different bodies.
Four sources of mismatch
1. Morphology
Joint count, link lengths, gripper geometry, balance and workspace determine which motions are physically possible.
2. Sensors
Camera placement, resolution, depth sensing, tactile feedback and proprioception change the observations available to a policy.
3. Action spaces
One platform may expose joint targets while another uses Cartesian poses, velocities or learned low-level controllers.
4. Dynamics
Mass, friction, actuator response, compliance and contact dynamics alter the outcome of the same nominal command.
How researchers reduce the gap
- Normalize observations and actions into shared representations.
- Condition models on embodiment metadata.
- Retarget demonstrations between kinematic structures.
- Train on multiple embodiments instead of a single robot.
- Use simulation to vary morphology and dynamics.
- Fine-tune on a small amount of target-robot data.
Why the gap matters for robot foundation models
A foundation model for physical intelligence needs more than broad visual or language knowledge. It must understand how goals become actions under the constraints of a particular body. Closing the embodiment gap is therefore a central problem in scalable robot learning.
Related topics
See What Is Robot Embodiment?, Cross-Embodiment Robotics and Robot Foundation Models.