Knowledge Transfer Between Different UAVs for Trajectory Tracking

Robots are usually programmed for particular tasks with a considerable amount of hand-crafted tuning work. Whenever a new robot with different dynamics is presented, the well-designed control algorithms for the robot usually have to be retuned to guarantee good performance. It remains challenging to directly program a robot to automatically learn from the experiences gathered by other dynamically different robots. With such a motivation, this letter proposes a learning algorithm that enables a quadrotor unmanned aerial vehicle (UAV) to automatically improve its tracking performance by learning from the tracking errors made by other UAVs with different dynamics. This learning algorithm utilizes the relationship between the dynamics of different UAVs, named the target and training UAVs, respectively. The learning signal is generated by the learning algorithm and then added to the feedforward loop of the target UAV, which does not affect the closed-loop stability. The learning convergence can be guaranteed by the design of a learning filter. With the proposed learning algorithm, the target UAV can improve its tracking performance by learning from the training UAV without baseline controller modifications. Both numerical studies and experimental tests are conducted to validate the effectiveness of the proposed learning algorithm.