Adaptive Real-time Nonlinear Model Predictive Motion Control

In this paper we present a framework for realtime, full state feedback, nonlinear model predictive motion control of autonomous robots. The proposed approach uses an iterative optimization algorithm, namely iterative Linear Quadratic Gaussian (iLQG) to solve the underlying nonlinear optimal control problem, simultaneously deriving feedforward and feedback terms. The resulting motion controller is updated online by continuously rerunning the solver in a model predictive control (MPC) setting. An additional optimization loop around the optimal control algorithm allows for a realtime, situation-dependent adaptation of the solver’s parameters. This adds the possibility to influence the high level behavior of the system such as adapting the controllers time horizon or cost function. The performance of the proposed approach is validated in simulation and on the ball balancing robot Rezero. Therefore, this work presents one of very few implementations of full-state feedback, nonlinear, model-predictive control for motion control on a real robot. Results show that the framework is able to produce an optimized behavior of the system that is robust to large disturbances. The efficient implementation allows us to run the framework at high frame rate in real-time on standard computer hardware.