f76a89f0cb91bc419542ce9fa43902dc-AuthorFeedback.pdf

We'd like to first thank the reviewers for their constructive feedback. A: The human motion synthesis experiments9 are designed specifically to show the generalization of RFC. Note that during test time the RFC policy willnotbe10 further trained to imitate the kinematic output, and it is basically tasked to do one-shot imitation. At test time, the11 motions generated bythe kinematic policyare different from training (different data) and asinput tothe state ofthe12 RFC policy they are analogous to the goals in DeepMimic. Without the residual forces, the agent often falls to the13 ground as demonstrated in the video (please see posing and walking dog motions starting from 3:31).