Non-Prehensile Manipulation Learning through Self-Supervision

摘要

Manipulation is one of most emerging research and development areas in the field of robotics. Recently, state representation learning for control has been gaining attention. In this paper, we proposed a novel learning model based on neural networks in order to sample the actions of the robot to push objects to desired positions. Furthermore, an intuitive method was proposed to enable the robot to collect training data in an efficiently way. Specifically, a fully convolutional network encodes observations into latent space, and a mixture density network is implemented to infer an action distribution, since there are an infinite number of possible actions that may result in the same change of the state of the object. Through extensive experimental simulations and comparisons with the existing models, we demonstrated the efficiency of the proposed method applied to non-prehensile manipulation, such as pushing or rotating of small objects on the table.