Author:

Liu, Yixuan
Supervisor:Prof. Gudrun Klinker
Advisor:Eichhorn, Christian (@ga73wuj)
Submission Date:[created]

Abstract

Seabirds like albatrosses are able to fly long distances over oceans for hours and even for days without flapping their wings. They can travel in such an efficient way because they obtain energy from shear wind. This flying behavior pattern is called dynamic soaring. Researchers can apply this pattern to unmanned glider drones in an autonomous way. The trajectories are either calculated on-time or off-time. In other words, they are computed on the glider while flying or on a workstation before take-off. This project shows one of the on-time approaches, deep reinforcement learning-based autonomous dynamic soaring. A model is trained first and can then be used on-time during flights. This work mainly contributes to minimize the neural network training time with an end-to-end GPU acceleration, and how the convergence in training improves the trajectory planning performance and yields good flight guidance results.

Results/Implementation/Project Description

Conclusion

[ PDF (optional) ] 

[ Slides Kickoff/Final (optional)]