쿼드로터 동역학 대칭을 이용한 고장 감내 제어 정책 학습 방법

Abstract

Drones are widely used in many domains in recent decades due to their agility, versatility, and ease of control. However, because of nonlinear dynamics, their control strategy is complicated. Especially when the drone has lost the sensor signals or action inputs, it can suffer from bad control, which can lead to crash at the worst cases. To address this issue, reinforcement learning (RL) agent have been widely researched to work as a controller. RL can effectively learn how to control the quadrotor even with motor failure using a lot of data. For the possibility of fault of any motors, I propose to use the symmetry of the dynamics of quadrotors. Using the configuration and symmetry of quadrotors, the data interacted for the fault of one specific motor can be augmented to the different motors. This approach showed faster convergence and final performance. This can be used for sample-efficient learning and safe learning of the quadrotors in real environment.