Many position-controlled robots are being used in research and industry in the world, but many tasks require torque control instead of position control, in order to exert specific forces in the environment. This is often called the admittance control problem. In this paper, we present a solution for position-controlled robots by estimating their hidden internal control law using Neural Networks and mitigating the fitting errors with an integral term in the control law. Compared to classical approaches, we no longer consider that the control law is decoupled between motors but it can be highly sophisticated and nonlinear. We show our results in simulation by performing torque tracking and force-position task control.