This research investigates and proposes a new method for obstacle detection and avoidance on quadrotors. One that does not require the addition of any sensors, but relies solely on measurements from the accelerometer and rotor controllers. The detection of obstacles is based on the principle that the airflow around a quadrotor changes when the quadrotor is flying near a surface. A well-known example of this is the ground effect, an increase in lift force close to a ground surface. Similarly, a change in dynamics occurs when a quadrotor is flying close to a wall or ceiling. The proposed method uses a reinforcement learning controller to detect obstacles based on these measurements, and take action to lead the quadrotor back to safety. A proof-of-concept of this method is developed by training a reinforcement learning agent to avoid obstacles beneath a descending quadrotor. This is first done in a simulated environment, where the influence of hyperparameters, the amount of noise in the state signal, and the number of training episodes are investigated. The best performing agent from simulation is evaluated during a flight experiment with the Parrot Bebop 1 drone, where it is able to prevent the quadrotor from hitting the obstacle in 80% of the episodes. Furthermore, it is shown that the same level of performance can be achieved, by learning fully from scratch, in-flight, without prior knowledge or training, during 50 real flight training episodes. An approach for extending this method to the avoidance of walls, ceilings, and smaller obstacles is discussed and the expected performance when doing so on the Parrot Bebop 1 is assessed. Additionally, it is shown that this method can easily be extended to other quadrotors, by demonstrating this on the Parrot Bebop 2 quadrotor.