Current whole-body control methods have incomplete embodiment representations which ignore the graph connectivity of the robot, and thus cannot generalize well to an embodiment with a varied graph structure.
This paper propose Graph Embodiment Transfoer, an encoder that uses the embodiment graph as a structural bias in network architecture. It consists of nodes representing local joint information, directed edges representing links connecting parent and child joints, and undirected edges between joints.
They also propose a parent-child encoding lets the model modulate attention based on directed connectivity of the forward joint chain.