We propose a new method for object pose estimation using a force-torque sensor and deep neural network. To estimate the pose of object, our method uses contact positions and force directions obtained by a force-torque sensor attached on a robot end-effector. In addition, sets of the contact data have characteristics of both various length and random ordered. For the purpose to deal with such data sets, which conventional neural networks cannot, we employ a symmetric function. The proposed method can estimate 6 degrees of freedom pose of objects without considering optical properties. We confirmed in the simulation that our method can estimate the pose for different types of objects: nut, bolt, and wrench. The experiment of pose estimation for nut in a real environment shows that the performance of 3.9mm in position error, 12.7 degrees in angular error, and 20,000 times faster computation time than the conventional method.