Unreal Engine with Airsim API.
This work demonstrates 5 learning episodes of an intelligent drone which starts with no information of its environment and learns to find the orange balls in the given city domain.
Briefly, the proposed framework can incentivize the operating agent with rewards collected from the drone’s camera (object detection), and punish based on the power use at each step. The former policy enables exploration of large areas where no distribution model for the target object is in hand, while the latter allows traveling for long distances in wind direction with minimal power use.
For the technical details on the framework and the implemented reinforcement learning algorithm please refer to:
[ Ссылка ]
