Abstract
This paper presents a novel computational approach for modeling and generating object manipulation behaviors by a humanoid robot. A time-delay deep neural network is applied for modeling the multiple behavior patterns represented with multi-dimensional visuomotor temporal sequences. Thanks to the efficient optimization capability of the Hessian-free deep learning algorithm, the proposed mechanism is successfully trained with six different behavior patterns without any dedicated feature extraction preprocessing. The experimental results show that the proposed mechanism can perform as a temporal sequence predictor as well as a cross-modal memory retriever. The result from a robot control experiment in the real environment demonstrates that the proposed mechanism can be utilized for adaptively selecting behaviors in accordance with the consecutive environmental changes.
