In the field of robotics, model predictive control is considered as a promising control strategy due to its inherent ability to handle nonlinear systems with multi-dimensional state spaces and constraints. However, in practice, the implementation of model predictive control for a nonlinear system is not easy, because it is difficult to form an accurate mathematical model for a complex nonlinear system. Moreover, the time required for solving a nonlinear optimization problem depends on the complexity of the system and may not be suitable for real-time implementation. In this thesis, a general approach for implementing model predictive control for nonlinear systems is proposed, where a physics-based simulator is used for the prediction of the states and a stochastic optimization based on particle belief propagation is used to solve the optimization problem. To study the ability of the controller, a nonlinear robotic system is built. The designed controller is capable of handling nonlinear system for both single variable and multiple variables. For the current system, the controller is unable to solve the optimization problem in real time with the presence of constraints. The proposed method provides a simpler approach for implementing model predictive control, which can be used for a range of robotic applications. However, in this method, the capability of the controller depends on the physics engine's ability to simulate different physical systems and the speed and accuracy of the physics engine.