Stochastic Multiple Shooting Trajectory Optimization via Sequential Local Policy Evaluation

The paper proposes a stochastic multiple-shooting trajectory optimizer that divides long control horizons into short action sequences connected by local feedback policies. Its goal is to improve sample efficiency and convergence to terminal sets compared with stochastic single-shooting methods such as MPPI. The authors also describe estimating approximate system Jacobians entirely from rollouts, allowing use with black-box or learned dynamics. According to the abstract, the method is evaluated on analytical and neural-dynamics cartpole swing-up tasks and a VTOL quadplane executing a high-angle-of-attack precision post-stall landing maneuver.