| A straightforward approach to the calibration problem for sensor-guided assembly systems is proposed. In this approach, called direct calibration, the entire assembly system--the robot, the sensors, and the parts to be assembled--is calibrated for the required assembly task in a single procedure to directly determine the relationship between the part feature information in sensor coordinates and the part location (position and orientation) in robot coordinates. The direct calibration method consists of these three steps: calibration data are generated by using the robot to move the part(s) to be assembled under the view of the sensors; the best-fit mapping representing this assembly process is calculated; and this mapping is used in production to estimate the part locations from the current sensor data. The key features of direct calibration are: accurate CAD models of the parts are not required; robot kinematic errors, geometric errors and joint backlash are accommodated; sensor/part interaction non-idealities are accommodated; and it is simple to use and fast to setup on the factory floor.;Three different methods are used for learning the mapping represented by the input-output data gathered at calibration. The first two methods, linear regression with quadratic basis functions and feed-forward neural networks, are generic mapping methods in that they only consider the calibration data and no prior knowledge of the system is required. The third method, parameter identification, is based on a general physical model of the system with the unknown model parameters to be determined from the calibration data.;This dissertation presents a comprehensive formulation of the direct calibration technique for sensor-guided assembly, extends prior 3 degree of freedom (3-DOF) work with simple part geometries to the general 6-DOF problem with complex part geometries, and develops new model-based parameter identification techniques for learning the process mapping. Simulation results and empirical data are presented that demonstrate the feasibility of direct calibration techniques using structured-light sensors for the automated insertion of automotive windshields, and the flexible feeding of complex injection-molded computer printer parts. |
