Recently, a robotic percussive riveting system has been developed at Ryerson University for an automation of percussive riveting process of aero-structural fastening assembly. The system consists of a robot holding a percussive riveting gun equipped with a rivet feeder, a gantry holding a working panel of aero-structure, and a position visual sensor. Prior to riveting, the robot is required to first position and then insert a rivet precisely into a hole on the panel without engaging the panel to prevent potential damage. The underlying challenges to precise insertion are various sources of system uncertainties, mainly including alignment errors among coordinate systems of the robot, panel and sensor, and relatively poor absolute positioning accuracy of the robot due to mechanical deflection, assembly clearance, and machining tolerance. For this reason, the research of relative pose estimation between the robot and panel has been carried out pertaining to these challenges.

Essentially, pose estimation is proposed for robotic percussive riveting, which estimates the relative pose between two rigid bodies based on noisy visual measurements of point features on rigid bodies. Three categories can be classified, namely, static, dynamic, and robust pose estimation. Firstly, static pose estimation is the parameter estimation of static relative pose transformations among a number of frames, which solves the issue of alignment errors. Direct solutions of static relative pose estimation are derived based on least-square methods. Secondly, to tackle the issue of poor absolute positioning accuracy of the robot, dynamic relative pose estimation is proposed addressing a state estimation of relative poses during motion. Iterative extended Kalman filter method is adapted for the state estimation. Thirdly, for robustness against outliers of point measurements, robust pose estimation is proposed based on an outlier diagnosis using the technique of relaxation of rigid body constraints. Indeed, outlier diagnosis is a pre-processing of point measurements, in which outliers are detected and removed prior to the relative pose estimation. Further, a decorrelation method is proposed for measurement calibration using multivariate statistical analysis to find an optimal sensorto-target configuration. As a result, each coordinate measurement is close to uncorrelated and it allows for a simple calibration.