Neural Network Feedforward Control for Pneumatic Hexapod Excavator Simulator

The dependency on accurate parameters is still one of the main problems for the model-based control of robots. If such accurate parameters are hard to acquire, then often traditional PID controllers are used instead, even if the control performance is barely sufficient. This paper proposes the combination of a standard PID controller with a data-driven feedforward control for the position control of a Stewart platform, used as excavator simulator and powered by fluidic air muscles. In detail, a neural network was trained with static robot poses as input to compensate the nonlinear effects of a centralized spiral spring with unknown nonlinear characteristics. The feedforward control via neural network was implemented on the robot hardware and tested on two different trajectories. The proposed control strategy leads to a superior control performance compared to the standard PID controller. In contrast to model-based control approaches, there is no need of knowing any parameters of the robot model.