In imitation learning, it is critical that the imitator estimate not only the trajectory of the demonstrator but also its itentions. This is particularly true when the physical parameters between them are different. In thi paper we propose a new approach for imitation learning. The imitator first estimate the subgoal of the demonstrator from its trajectory, and choose it to become its own subgoal. This way, the imitator evaluates its own behaviour rather than simply copying the output of the demonstrator, and can account for differences in dynamics in a natural way.
展开▼
