Direct Coupling of Multisensor Information and Actions for Mobile Robot Behavior Acquisition

Many conventional methods for multisensor fusion have been based on the prede ned selection such as \which sensor is used for which situation," and mainly they have seemed to concentrated on the reconstruction of a 3-D geometric model of the environment supposing such a model would be generally useful. However, optimality of a priori selection by the programmer is not guaranteed, nor is such a geometric model easy to be applied. This paper proposes a method which enables a mobile robot to acquire a purposive behavior for accomplishing a given task by directly coupling multisensor information and actions through interaction between the robot and its environment. We use reinforcement-learning scheme to formalize such a coupling process. First, we de ne states described by combinations of various kinds of data provided by di erent types of sensors and motor commands to the mobile robot. Then, we acquire pairs of robot actions and states suitable for achieving the given goal by using Q-learning algorithm. As a result of learning, the goal-directed behavior is obtained and information needed for the current subtask is automatically selected among multisensor information. The validity of the method is demonstrated by computer simulations and real robot experiments.