Abstract
In this paper, an action space description“StateNet”is introduced. It includes an automatic error recovery function. Robots should be able to go on moving without human support even if they fall into an unexpected situation. However, it is very difficult for developers to write all error handlers, especially in the case the robot has various actions and acts in various situations. So the description must includes relations between actions and situations. The action space is described as a state transition graph which is defined in a state space constructed by sensory information. Using a state and path generation, a distance calculation and a path finding, the automatic error recovery function is realized. Finally, StateNet is implemented on a remote-brained humanoid robot and its effect is confirmed.
