A Safe And Efficient Parameter Optimization Method For Quadrotor: Onboard Implementation And Experiment

When we designing a controller for a quadrotor, the choice of parameters is a fundamental and important task.To achieve better control performance, manual tuning of parameters in the engineering practice is often required, which would be time consuming and laborious. Therefore, there is an urgent need for a safe and efficient automatic parameter tuning method.In fact, the tuning of control parameters for a quadrotor is a classic EE (Exploration and Exploitation) problem, aiming to extend the safety bounds while finding a potentially optimal value. Many machine learning methods, such as deep learning, have been proposed to solve this problem, but these methods often fail to ensure the safety and stability of the closed loop system when performing parameter optimisation. Recently, Felix Berkenkamp proposed a safe Bayesian optimisation algorithm called SAFEOPT,which estimates the performance function of the UAV online via a Gaussian process, and ensures that the UAV automatically explores the optimal control parameters within a determinate safety threshold. When dealing with the trade off between exploration and exploitation in the optimisation of parameters, The original safeopt algorithm selects the point where Gaussian predictions are most uncertain by expanding the security frontier and finding the potential optimum. The algorithm can guarantee to find the optimum and ensure safety simultaneously.However, we find that the SAFEOPT algorithm consumes more time in the exploration of the safety boundaries, but we prefer to find the optimal parameters as fast as possible without considering the boundaries of the defined safety of the system. In order to improve the efficiency of the algorithm,we introduces a is an element of-decreasing algorithm, which improves the search rules of the original SAFEOPT. Initially, the algorithm is more focused on exploration to avoid falling into a local optimum, and when the number of explorations reaches a certain value, the algorithm pays more attention to exploitation, which is conducive to finding the optimal value quickly. Experimental results on a quadrotor show that the improved SAFEOPT method is able to find the optimal value more efficiently while ensuring the safety of the system.