Current quadcopter stabilization is done using classical PID controllers. They usually perform well expect for:
- altitude control, due to complex airflow interactions present in the system.
- when non-linearities are introduced, which is the case in clustered environments. We envision reinforcement learning as a major breakthrough in the field of robotic control as it eliminates the need for a predefined controller structure which limits the overall performance and costs more human effort. In other words, reinforcement learning avoids creating manually-tuned control algorithms to complete specific tasks. This technique could lead to building tailored, adaptable, autonomously learned controllers for drones and other types of robotic technologies.
The source code can be found inside the code folder. The final version of the code we used can be found under code/FVI/
The other durectories contain the following:
- The FVI folder contains an implementation where the control vector matches what would be used in a real life drone.
- The FVI/acc folder contains a version of the simulator in which we only control the spatial acceletarion of the drone (not really controlling thrust, roll pitch and yaw)
- The Matlab Simulator folder contains the initial implementation of the quadcopter dynamics
- The OpenAI folder contains some test we ran when attempting to use OpenAI gym.
- The gym folder was forked from the OpenAI repository (https://github.com/openai/gym) when we attempted to create a virtual environment for our problem.
- The value_it folver contains our initial implementation of the Fitted Value Iteration algorithm using different feature mappings and reward functions.