In order to solve the control problem of stochastic nonlinear robot systems, an improved RRT * algorithm with two-stage risk aversion architecture is proposed. The nonlinear dynamics is introduced by solving the NLP, and the approximate state distribution and distribution-robust collision detection are used to quantify the risk, so as to complete the path planning. Three kinds of controllers are demonstrated: LQR with linear dynamics around the reference trajectory, LQR with robust multiplicative noise term, and nonlinear model predictive controller. Laplace noise with heavy-tailed characteristics is imposed in the unicycle dynamics model. The results show that the algorithm is feasible.