Perception-Driven Motion Control Based on Stochastic Nonlinear Model Predictive Controllers

In mobile robotics, constraints may represent mobility issues, vision uncertainties, or localization uncertainties. In model predictive control (MPC) theory, constraint satisfaction is typically guaranteed through the use of accurate prediction models or robust control. However, although MPC offers a certain degree of robustness to system uncertainties, its deterministic formulation typically renders it inherently inadequate for systematically dealing with uncertainties. Furthermore, the model used by robust controllers is often not updated online. Towards this direction, this paper presents a stochastic nonlinear model predictive control (SNMPC) algorithm for active target tracking. Our goal is to use a SNMPC to penalize the undesired behavior, allowing the robot to converge to the optimal pose in order to observe the target optimally. We perform this by proposing uncertainty prediction models for both the target position estimation and the robot pose estimation. Based on these models, the predictive controller allows the robot to drive at or near its capabilities while respecting the imposed constraints. This paper presents real robot experiments in which the stochastic nonlinear controller provides satisfactory target tracking control.