Large and sudden changes in the torques of the motors of a robot are highly undesirable and should be avoided during robot control as they may result in unpredictable behaviours. One cause of large changes in torques is the presence of discontinuities in the constraints that the robot must satisfy, such as the avoidance of an obstacle or the breaking of contacts with the environment. In this paper, a model predictive control (MPC) approach to approximate constraints that can be predicted over a finite horizon is proposed to minimize the derivative of torques during robot control. The proposed method does not directly modify the desired task trajectory but the constraints to ensure that the worst case of changes in torques is well-managed. From the simulation results on the control of a Kuka LWR robot, it is shown that our approach significantly decreases the maximum derivative of joint torques for both force and acceleration task control examples.