Download MOV filei10MBj
Download MPG filei53MBj
A robot can learn to walk just by exploring smartly a sensorimotor space. In this experiment by Frederic Kaplan and Pierre-Yves Oudeyer, a progress-driven algorithm controls the movement of the different motors of a four-legged robot. For each motor, it chooses the period, the phase and the amplitude of a sinusoidal signal. The prediction system tries to predict the effect of the different sets of parameters in the way the image captured by a camera placed on the robots head is modi?ed. This indirectly re?ects the movement of its torso. At each iteration the algorithm produces the values for the next parameter set in order to maximize its expected progress in prediction. This means it practices the movements that are more likely to make it learns. After three hours of this curiosity-driven exploration, the robot masters a set of versatile walking skills. It learned how to walk although it was never directly rewarded for walking.