Enhancing the Precision of a Hydraulic Robotic Arm

Sammanfattning

This thesis explored the use of a Gated Recurrent Unit (GRU) based neural network, used as a feedforward controller in the electro-hydraulic driven robotic arm of DeLavals Voluntary Milking System (VMS). The research focused on developing a GRU-based feedforward controller and comparing its performance to the existing controller of the VMS in terms of velocity tracking accuracy, as the non-linear dynamics of the electro-hydraulic actuators pose a challenge for traditional methods of system modelling. The network was trained on data collected from the robotic arm’s movements, enabling it to learn the complex relationship between desired velocities and the corresponding outputs. This trained network was implemented as a feedforward controller, directly generating actuator commands to achieve the desired velocity trajectory. The performance comparison utilises Mean Squared Error (MSE) to evaluate the tracking accuracy of both controllers. The results demonstrate that the GRU-based feedforward controller significantly outperforms the existing controller, achieving a 53% lower MSE value. These results further signify the potential of GRU neural networks for improved control of robotic manipulators with complex dynamics.