Co-imitation: Learning Design and Behaviour by Imitation

The co-adaptation of robots has been a long-standing research endeavour with the goal adapting both body and behaviour robot for given task, inspired by natural evolution animals. Co-adaptation potential to eliminate costly manual hardware engineering as well improve performance systems. standard approach is use reward function optimizing morphology. However, defining constructing such functions notoriously difficult often significant effort. This paper introduces new viewpoint on problem, which we call co-imitation: finding morphology policy that allow an imitator closely match demonstrator. To this end propose co-imitation methodology matching state-distributions Specifically, focus challenging scenario mismatched state- action-spaces between agents. We find increases similarity across variety tasks settings, demonstrate transferring human walking, jogging kicking skills onto simulated humanoid.