Utilizing online learning based on echo-state networks for the control of a hydraulic excavator

In this paper, we present a control framework for the control of a hydraulic excavator. An excavator can be viewed as a robotic manipulator that interacts with the environment. It follows that the control method employed to control the excavator must take into account the complex soil–tool interaction in order to achieve the desired trajectory of the manipulator. Impedance control has been proven to be effective in this aspect in that it provides an unified approach to constrained and unconstrained motion. Another important aspect when considering the automation of an excavator is the control of the hydraulic servo system. Obtaining a useful explicit model for the control of hydraulic servo systems is not a simple task due to their inherent complex nonlinearity. Therefore, control techniques that do not require an explicit representation of the plant are required. In this work, we integrate two controllers for the automation of an excavator. To control the rigid-body motion of the excavator, impedance control and sliding mode control are applied. The results are desired cylinder forces that are required to achieve the desired trajectory. Given the desired cylinder forces, an online learning control method is employed to control the hydraulic servo system so that the desired forces are generated. Echo-state networks, which are a class of recurrent neural networks, are utilized within the online learning control framework in order to learn an inverse model of the hydraulic servo system. Thus, the online learning control framework does not require an explicit model of the plant and also adapts to the plant using only the input and output signals. We present results of the proposed control framework on an excavator simulation environment that has been verified based on operation data from a real hydraulic excavator.