Flight test of Quadcopter Guidance with Vision-Based Reinforcement Learning

Reinforcement Learning (RL) has been applied to teach quadcopters guidance tasks. Most applications rely on position information from an absolute reference system such as Global Positioning System (GPS). The dependence on “absolute position” information is a general limitation in the autonomous flight of Unmanned Aerial Vehicles (UAVs). Environments that have weak or no GPS signals are difficult to traverse for them. Instead of using absolute position, it is possible to sense the environment and the information contained within it in order to come up with a “relative” description of the UAV’s position. This paper presents the design of an RL agent with relative vision-based states and rewards for the teaching of a guidance task to a quadcopter. The agent is taught the task of turning towards a red marker and approaching it in simulation and in flight tests. A more complex task of travelling between a blue and a red marker is trained in simulation. This work shows that relative vision-based states and rewards can be used with RL to teach quadcopters simple guidance tasks. The performance of the trained agent is inconsistent in simulation and flight test due to the inherent partial observability in the relative description of the state.