A humanoid behavior modification system by monitoring & evaluating causality between sensors & actions

The idea of automatic behavior modification that runs in parallel with the main behavior generation system is important in order to adapt to many unexpected changes that robots are exposed. Against some of the unexpected changes, robots have to decide whether they stop doing the current task and search for the changes or ignore the changes. By monitoring causality between their sensors & actions while doing the current behavior, robots can expect which sensors are to change in the current behavior. Therefore, they can evaluate how unexpected the changes are. This paper proposes a method to modify robots´ behaviors automatically when the unexpected sensor changes happened. For realizing such system, we developed the mechanism for learning the connection strength between sensors and actions by monitoring their synchronous changes during current behavior & evaluating how unexpected the changes of each sensors are. Furthermore, simple behavior modification experiments using real humanoid are shown for the confirmation of the feasibility of the proposed system