P3 and ASIMO - The Honda Humanoid Robots

Product Name:

Product Description

For a biological organism or an artificial agent like a robot, the ability to control its limbs and actuators is very important. This kind of control is called motor control. From an engineering perspective, control theory provides the means to determine motor controllers analytically. But this works only for robot setups in well-defined environments and restricted task domains. For biological modelling or for the design of autonomous artificial agents (robots), adaptive motor learning becomes as important as motor control itself. Motor controller are acquired and changed in the interaction of the organism with its environment. For example, the shape of the human body changes throughout lifetime, making it necessary to change the forces which are exerted by the muscles for the accomplishment of certain tasks. Furthermore, humans and animals are able to learn new motor tasks and to improve their physical skills. For saccade learning, we developed a motor learning strategy called \"learning by averaging\" (LBA). Saccades are fast fixation movements of the eyes which we simulate with our robot camera head. As saccade controller, we use a multi-layer perceptron (MLP). The basic idea of LBA is to search at random in the neighbourhood of the output of the MLP in motor space for saccades which are slightly better than the saccade produced by the controller network, and which bring the target object closer to the center in both camera images. These improved saccades are used as learning examples for network adaptation. In the process of learning, over- and undershoot saccades cancel each other out, resulting in more precise motor output of the controller network. This \"cancelling out\" only works because the MLP as function approximator adapts to the average of the over- and undershoot saccades.

Company Details

Based at the University of Bielefeld. Main research topics include biorobotics, cognitive robotics, sensorimotor control, and parallel computation. more