Expressing the full-body character as a single rigid body, we can obtain a control policy that can adapt to various unobserved environmental changes and controller transitions. We demonstrate the ability of the policy with environments such as running on uneven terrain or pushing a box, and transitions between learned policies.