A Novel Reinforcement Learning-Based Robust Control Strategy for a Quadrotor.

In this article, a novel reinforcement learning (RL)-based robust control approach is proposed for quadrotors, which guarantees efficient learning and satisfactory tracking performance by simultaneously evaluating the RL and the baseline method in training. Different from existing works, the key novelty is to design a practice-reliable RL control framework for quadrotors in a two-part cooperative manner. In the first part, based on the hierarchical property, a new robust integral of the signum of the error (RISE) design is proposed to ensure asymptotic convergence, which includes the nonlinear and the disturbance rejection terms. In the second part, a one-actor-dual-critic (OADC) learning framework is proposed, where the designed switching logic in the first part works as a benchmark to guide the learning. Specifically, the two critics independently evaluate the RL policy and the switching logic simultaneously, which are utilized for policy update, only when both are positive, corresponding to the remarkable actor-better exploration actions. The asymptotic RISE controller, together with the two critics in OADC learning framework, guarantees accurate judgment on every exploration. On this basis, the satisfactory performance of the RL policy is guaranteed by the actor-better exploration based learning while the chattering problem arisen from the switching logic is addressed completely. Plenty of comparative experimental tests are presented to illustrate the superior performance of the proposed RL controller in terms of tracking accuracy and robustness. [ABSTRACT FROM AUTHOR]