|
In recent years, neural-network-based adaptive dynamic models are commonly used to estimate and control flight dynamics for drones and multi-copters. However, most of them simply use networks to optimize few parameters in control policy. There are still large improvements on the structural layout for robust model-free control applications. Therefore, to achieve a similar intelligence of human is still challenging due to their difference in basic mechanisms and difficulty in network modeling. In this paper, we design a control learning algorithm which combines reinforcement learning with neural networks simplified from human cerebellar motor learning model. The algorithm learns parameters by statistically measuring the performance and analyzing the input-output relationship on real-time episodes. In local linear systems, parameters are learned with respect to a spatial function of environment state and subjective expectation. Compared with other methods using static models, the most obvious advantage of this algorithm is that it can learn complex dynamics of alternative degrees of freedom while the dynamics are difficult to be formulated by equation set. Besides, the algorithm is suitable for individual systems, without prior knowledge about system geometry, centre of gravity as well as installation error, since it learns the dynamic effects directly relevant to the optimal guidance and control behavior in unknown or partially known environments instead. Experiment verifies the algorithm in a practical way. In the experiment, the algorithm is implemented to a quad-copter and it can learn the flight control policy from zero-state without any prior knowledge. The flight quality is tested to be equal to accurate control model at outdoor flight experiences. By repeated experiment, the algorithm is demonstrated to have good robustness to control different physical models and the potential to explore alternative dimensionality.
|
