Deep Spatial Autoencoders for Visuomotor Learning

摘要: Reinforcement learning provides a powerful and flexible framework for automated acquisition of robotic motion skills. However, applying reinforcement requires sufficiently detailed representation the state, including configuration task-relevant objects. We present an approach that automates state-space construction by state directly from camera images. Our method uses deep spatial autoencoder to acquire set feature points describe environment current task, such as positions objects, then learns skill with these using efficient based on local linear models. The resulting controller reacts continuously learned points, allowing robot dynamically manipulate objects in world closed-loop control. demonstrate our PR2 tasks include pushing free-standing toy block, picking up bag rice spatula, hanging loop rope hook at various positions. In each automatically track their robot's arm.