A model-based learning controller with predictor augmentation for non-stationary conditions and time delay in water shooting

摘要

As evident in the Fukushima incident, shooting water to hit a target at a distance under a windy condition from a moving vehicle is challenging. The challenges include (i) the non-stationary condition caused by the wind and (ii) the time delay due to the distance between the water source and target. This paper proposes a model-based learning controller to address these issues. The proposed controller can adapt the water's shooting angle to different wind conditions or vehicle motions by measuring only the shooting error and angle. First, a forward model uses the current shooting error and its corresponding past shooting angle to predict the future shooting error, thereby addressing the time delay issue like the Smith Predictor. Next, the current shooting angle and its predicted future error are given to an inverse model as the augmented predictor to determine the required shooting angle necessary to achieve zero error. Both the forward and inverse models are learned using the Receptive Field-Weighted regression (RFWR) algorithm. Interpolation, cross correlation and active probing based techniques are developed to estimate the time delay adjustment needed to synchronize the shooting angle and error feedback for model learning. Experimental results obtained from computer simulations indicate that the proposed controller can adapt to non-stationary conditions and address the time-delay issue. The performance of the controller outperforms human operators and a simple PID controller in water-shooting tasks under changing wind and vehicle motion.