A Simple Neural Network for Efficient Real-time Generation of Dynamically-Feasible Quadrotor Trajectories

In this work, we study the problem of efficiently generating dynamically-feasible trajectories for quadrotors in real-time. A supervised learning approach is used to train a simple neural network with two hidden layers. The training data is generated from a well-established trajectory generation method for quadrotors that minimizes jerk given a fixed time interval. More than a million dynamically-feasible trajectories between two random points in the three-dimensional (3D) space are generated and used as training data. The input of the neural network is a vector composed of initial and desired states, along with the final trajectory time. The output of the neural network generates the motion primitives of the trajectories, as well as the duration or final time of a segment. Simulation results show extremely fast generation of dynamically-feasible trajectories by the proposed learning algorithm, which makes it suitable for real-time implementation.