| dc.description.abstract | This thesis is written for Heimdall Power AS to gather experiences and test the effectiveness of reinforcement learning as a solution for their drone project. Heimdall Power makes sensors mounted to the power grid. These sensors are mounted by linemen, and Heimdall Power is looking into having this process replaced by autonomous drones. Such a solution would have a pilot position the drone close to the power line, and then the drone would take over and autonomously position itself close above the power line such that the sensor can be mounted.
A Deep Q-Network, using a depth camera as input, has been trained inside of the simulator Airsim to position itself in the right orientation and distance to a power line such that a sensor could be mounted. The most successful agent reached a success rate of 86%. The agent was also tested on simulated weather and night time, with only small changes to the performance. However, The performance did drastically fall when tested on an environment different than the one trained on, such as an agent trained with only one power line visible tested on an environment with multiple visible power lines and vice versa.
This thesis has shown some success using reinforcement learning as a solution for Heimdall Powers drone project. The solution made in this thesis can not be used directly, but the experiences gathered around reinforcement learning and sensors will be of great help when the drone projects continue in the summer of 2020. | |