| Assembly operations stand to benefit significantly from the development of intelligent programming techniques to compensate for the intrinsic real-time uncertainties that have traditionally restricted their full automation. The limitations of robot positional accuracy, and the effects of sensorial and control uncertainties are among the factors that may hinder the achievement of a successful assembly, therefore causing the need for constant human supervision during the process. This ultimately contributes to the high costs involved with industrial assembly operations. In line with current research, this thesis investigates the application of novel control algorithms to robotic assembly, aiming to provide industrial manipulators with improved flexibility and adaptability by emulating human like sensori-action and cognitive behaviour. The work described in this thesis considers force driven peg in hole insertions as a fundamental primitive for automated assembly operations. The contact between mating components is analysed from a sensorial perspective, which enabled the identification of geometrically dependent and independent features in the force and torque signal retrieved from a wrist-mounted sensor. These findings lie at the foundation of the intelligent assembly controller proposed, which for the first time enables several sources of readily available information to be included in the control loop, without resorting to external supervision. These include a CAD model of the parts which is merged with the geometrically dependent component of the online sensed contact information using a purposely developed neural network technique. The correction of linear misalignments is therefore achieved by estimating the contact's location and selecting an optimal assembly path. The geometrically independent component of the force and torque signal is on the other hand involved with the correction of the angular misalignment between the components. A novel autonomous agent, namely the ART-R network, has been developed, to learn via reinforcement the manipulative skill required for the correction. Results from real-time insertions performed with a PUMA 761 industrial manipulator and wrist-mounted force and torque sensor show the ability of the controller to execute assembly operations starting with a combination of significant linear and angular misalignments. The experiments extend to consider various components cross-sections and direction of insertions, providing consistent indications on the validity of the approach. The achieved performance, paired with the notable reduction of external supervision and expert knowledge required during the execution of the task, represents a considerable improvement over previous approaches and a significant move towards the development of reliable autonomous robots for industrial assembly. |
