Force-Guided Robotic Assembly Process: Control and Contact-State Recognition

This thesis addresses developing novel Contact-State (CS) modeling, control strategy, and environment position localization (position searching) for force-guided robotic assembly processes of rigid and flexible objects. For the CS modeling, the wrench (Cartesian force and torque) signals of the manipulated object are captured for different phases of the considered assembly processes and using the Expectation Maximization-based Gaussian Mixtures Model (EM-GMM), a recognizer is developed for each CS of the assembly. The suggested EM-GMM CS modeling scheme is shown to have excellent Classification Success Rate (CSR) with reduced computational efforts. For the control part, it is shown throughout the thesis that a force-guided robotic assembly process is a hybrid nonlinear system with arbitrary switching signal resulted from the constraints arbitrary switching during the assembly. Furthermore, the robot dynamics is frequently unknown, which is the case in many industrial robots, that would make the force-guided robotic assembly process to be an unknown hybrid nonlinear system with arbitrary switching. In order to overcome such a control challenge, a Decentralized Robust Adaptive Fuzzy Control (DRAFC) strategy is derived that guarantees stable performance under constraints arbitrary switching and unknown dynamics. For the environment position localization, the EM-GMM CS modeling scheme is integrated with a spiral search path and the precise hole position is identified for cases of position uncertainty. Experiments are conducted on a KUKA Lightweight Robot (LWR) doing different force-guided assembly tasks for rigid and flexible objects. Excellent performance is reported for the proposed EM-GMM CS recognition scheme, the DRAFC strategy, and the suggested position searching algorithm. The suggested EM-GMM CS recognition, DRAFC strategy, and position localization schemes are compared with the availably corresponding schemes and the superiority of the suggested schemes is shown. The reasons behind the superiority of the EM-GMM CS recognition scheme are the accommodation of the captured signals non-stationary behavior, employing optimized number of GMM components in the modeling process, and employing the EM algorithm that iteratively increases the log-likelihood. The causes behind the superiority of the DRAFC strategy are addressing the unknown nonlinear dynamics of the robot, accommodating the constraints arbitrary switching, and the robustness against possible dynamics parameters drift. The reasons behind the surpassing of the suggested position localization strategy are the robustness against the surface roughness and reduced computational efforts. The proposed EM-GMM CS modeling scheme, DRAFC strategy, and position searching scheme are applied to the entire peg-in-hole assembly processes of rigid and flexible objects. Excellent Localization Success Rate (LSR) was resulted when using the suggested schemes. Furthermore, the proposed CS modeling scheme, control strategy, and localization approach are applied to a couple of applications in automotive industry; the first one is the camshaft caps assembly of a cylinder head and the other is the air-intake manifold assembly of a powertrain. Efficient force-guided robotic assembly processes are obtained for both considered applications.