Abstract:
One of the most desirable characteristics of UAVs is autonomous navigation, which
aids in a variety of applications such as search and rescue missions, delivery,
surveillance, and so on. UAVs are equipped with various sensors such as stereo
cameras and LIDAR to gather information about their surroundings for navigation
purposes. UAVs fly safely by using a 3D map and an obstacle-free path, but this
requires more on-board computational resources and power. Because UAVs lack these
resources due to their small size and limited battery capacity, using path planning
algorithms is not an efficient solution. To address this issue, we used deep
reinforcement learning to learn the action policy on its own and take actions based on
it in order to navigate successfully without the use of a map. As indoor environments
are more dynamic in nature that's why we have created multiple indoor simulated
environments in the Unreal Engine framework in which we have trained and tested
our drone.
