Model Predictive Local Motion Planning With Boundary State Constrained Primitives

Motion primitives are frequently used to find valid local trajectories for mobile robots, especially in cases where fast replanning is required, but the onboard computational power is limited. In this letter, we present a practical framework for constructing motion primitives from boundary state constraints, and then using them for online planning. The primitives are offline constructed with either a boundary value problem solver or a controller. They are then approximated with a neural network for fast evaluation during online optimization. The references and nominal inputs are generated in a receding horizon fashion by solving a model predictive control problem in the continuous domain with either gradient-based or gradient-free techniques. The proposed approach is computationally efficient and has been tested on quadrotors in real flight experiments, including sensor-based navigation, flying through a complex three-dimensional environment, dynamic obstacle avoidance, and tracking moving references.