Identification of reverse-action pairs using reversibility of actions

This paper presents a practical approach to identification of reverse-action pairs for reversibility-based learning to develop safe behaviors. The approach is to analyze reversibility of consecutive actions by introducing metrics defined on the actions and the states. The experiment is conducted in the Player/Stage simulator to test applicability of the approach. The robot is allowed to move freely in the environment by executing two-dimensional actions - moving forward/backward or rotating left/right. Reversibility of a sum of every consecutive pair of actions is then analyzed to identify the ones that undo each other. As a result, a set of action pairs are identified with a general rule “if moved/rotated by X, move/rotate by -X to undo”.