Trajectory Tracking via Probabilistic Movement Primitives for Hydraulic Manipulators

This paper introduces a novel application of Probabilistic Movement Primitives (ProMPs) for control of industrial hydraulic manipulators. We show how to include both manipulator kinematics and hydraulic cylinder dynamics in the trajectory learning. The solution is based on approximate quasi-static relation between hydraulic valve control commands and joint variables. The benefit of the approach compared to naively adding the control inputs in the ProMPs vector is that our approach reduces the required dimension of the vector and the corresponding cross-covariances. Moreover, this allows for the feedforward control commands to be sampled from ProMPs instead of being calculated during operation. Our results show that the proposed framework reduces the noise when compared to the latter. Trajectory learning is done using data collected from a real Komatsu PC138US-8 excavator, while demonstration of the efficacy of the control system is shown on a high-fidelity simulator of the machine. We also analyze different choices of the basis function and compare their suitability for encoding an excavator digging motion.