Location
Room # EB302
Start Date
5-6-2019 9:20 AM
Description
In robotics applications, Model Predictive Control (MPC) has been limited in the past to linear models and relatively short time horizons. In recent years however, research in optimization, optimal control, and simulation has enabled some forms of nonlinear model predictive control. The limiting factor for applying nonlinear MPC for robotics remains the computation necessary to solve the optimization, especially for complex systems and for long time horizons. This paper outlines and applies several methods in order to address the computational concerns related to nonlinear MPC. The dynamic model of the system to be controlled is approximated using a Deep Neural Network (DNN), then input trajectories over long time horizons are parameterized with a few parameters in order to decrease the optimization search space. Using the parameterized trajectories and the approximate model, an evolutionary algorithm is used to find optimal controls at the next time step. Simulations on torque limited robots performing a swing-up task demonstrate that the nonlinear Evolutionary MPC (NEMPC) is able to discover complex behaviors to accomplish the task. Comparisons with state-of-the-art nonlinear MPC algorithms highlight the fact that NEMPC does not require initialization with a candidate trajectory or policy in order to find a feasible (near optimal) trajectory.
Real-Time Nonlinear Model Predictive Control Using a Graphics Processing Unit
Room # EB302
In robotics applications, Model Predictive Control (MPC) has been limited in the past to linear models and relatively short time horizons. In recent years however, research in optimization, optimal control, and simulation has enabled some forms of nonlinear model predictive control. The limiting factor for applying nonlinear MPC for robotics remains the computation necessary to solve the optimization, especially for complex systems and for long time horizons. This paper outlines and applies several methods in order to address the computational concerns related to nonlinear MPC. The dynamic model of the system to be controlled is approximated using a Deep Neural Network (DNN), then input trajectories over long time horizons are parameterized with a few parameters in order to decrease the optimization search space. Using the parameterized trajectories and the approximate model, an evolutionary algorithm is used to find optimal controls at the next time step. Simulations on torque limited robots performing a swing-up task demonstrate that the nonlinear Evolutionary MPC (NEMPC) is able to discover complex behaviors to accomplish the task. Comparisons with state-of-the-art nonlinear MPC algorithms highlight the fact that NEMPC does not require initialization with a candidate trajectory or policy in order to find a feasible (near optimal) trajectory.
Comments
Session 2