We present a learning algorithm that realizes a simplegoal-reaching task for an autonomous vehicle when only visualinformation about the goal is provided. The robot has to reach a doorfrom every position of the environment. The state of the system is basedon visual information received by a TV camera placed on the mobilerobot. The vision algorithm is able to determine the relative positionbetween the vehicle and the door according to the slopes of the contourlines of the door. A learning phase is carried out in simulation toobtain the optimal state-action rules. The learned knowledge is thentransferred on the real robot for the testing phase. Experimentalresults show the generality of the knowledge learned as the real robotis always able to execute its paths towards the door in differentenvironments
展开▼
