Learning to land: autonomous quadcopter recovery from rotor loss using adaptive thrust vectoring

Quadcopter drones depend on four rotors to manage altitude and orientation. However, the failure of a single rotor compromises the drone’s ability to remain airborne and to land safely. This article introduces a solution enhancing attitude control in the event of a complete rotor failure by reducing roll, pitch, and yaw deviations during landing. This was achieved by actively tilting and panning the three remaining operational rotors. The controller for the pan and tilt mechanism was developed using the reinforcement learning approach. The solution was reached after the agent accumulated reward points over 4,000 training episodes when the feedforward thrust setting for the rotors was set to ‘48’. The uncontrollable attitude during flight was mitigated. The experiment was extended to explore the impact of changing the feedforward thrust setting to ‘60’. With the additional thrust, the unbalanced drone demonstrated better attitude responses during touchdown.

Actions (login required)

View Item