A machine learning approach to trajectory planning for unmanned aerial vehicles

Abstract

Towards this goal, we first create dynamic models of quadcopters and then construct typical optimization problems for time-optimal trajectory generation based on the proposed dynamic model. The formulated optimization problems are solved off-line for a range of initial and terminal flight states to generate waypoint-to-waypoint data-sets which contain all information about the optimization-based off-line trajectory planner.

Nowadays, unmanned aerial vehicles(UAVs) are commonly applied in people's daily life as well as in a variety of industries. There are still multiple challenges in achieving a more comprehensive application of UAVs, among which the challenge of achieving autonomous flight is the most important. Typical UAV autonomous systems consist of guidance, navigation, and control (GNC) systems, and trajectory planning is a critical component of all GNC systems. The general trajectory planning method is to generate a set of waypoints following which waypoint-to-waypoint trajectory planning for UAV's can be done. However, the computational cost can be so high that it will be hard to guarantee real-time trajectory generation. In order to increase the on-the-fly trajectory generation efficiency, the goal of this thesis is to propose a machine-learned trajectory planning algorithm for generating point to point time-optimal trajectory for a quadcopter.

Then we create the neural network structure and initialize the neural networks. After initialization, we train the neural network with the compressed new training data sets. Once the neural network is trained, the parameters of the neural network are updated, and a corresponding data-driven onboard path planner is created.

Next, we compress the information in the optimization-based trajectory data-sets by parameterizing them using a set of basis functions. These basis function parameters are saved in new data sets and separated into a training data set, a testing data set, and a validation data set for later use.

Finally, the neural network model is tested against out-of-sample data. The machine-learned trajectory generation model is checked for the approximation accuracy as well as any violation of the constraints enforced in the optimization problem; thus, this machine-learned model is proved accurate and feasible.