We present the first deep learning model to successfully learn control policies directly from high-dimensional sensory input using reinforcement learning.