Why cross-embodiment training helps

Robotics foundation models (e.g. trained on the Open X-Embodiment data) mix data from many different robot types. Why does this typically help?