Model Predictive Control with Torque Constraints for Velocity-Driven Robotic Manipulator

Abstract

Model Predictive Control is a well-known approach for executing the motion of a robotic manipulator along a certain trajectory. Most existing solutions consider the vector of joint angles as a state and the vector of joint torques as control input. However, some advanced collaborative manipulators do not allow direct torque control due to safety reasons. Available velocity control may lead to undesired inaccuracy and overdrives if actual torques are changing too fast. To avoid this, we propose a controller, which defines velocity control input, taking into account torque constraints. We develop a kinematic process model and a dynamic constraint model and insert them into the model predictive controller. Defining the realtime constraint function is a challenging task as the model of a robot's dynamic is nonlinear and computationally expensive. We propose a simplification approach to the constraint model and evaluate its applicability in numerical experiments. We evaluate our controller in simulation and with a real UR5 robotic manipulator. As an example task, we consider pushing the elevator button based on visual estimation of its coordinates. Experiments showed that the controller is able to provide smooth execution of this task.