Control based motion primitives for quadrotor trajectory generation

The local motion planning is critical for the safe operation of an autonomous vehicle such as the quadrotor in which motion primitives are usually adopted to find valid and efficient reaction. We present in this work a method that is capable of generating a class of motion primitives for the quadrotor. Compared to the previous approaches, these primitives are capable of minimizing multiple types of cost functions while satisfying the motion constraints of the quadrotor. Moreover, neural networks are adopted to approximate the generated motions to further boost the online evaluation efficiency. Finally, the local motion planning is achieved through solving a non-convex optimization problem with particle swarm techniques in a receding horizon fashion. The optimization process is a natural extension of the previous tree searching based methods. The proposed approach is computationally efficient and has been tested in a real environment on an actual quadrotor platform.