In the past, robotic assembly has required rigid fixturing and special purpose robotic tools for every assembly component. Unfortunately, rigid fixtures and special purpose robotic tools often have to be customized for varying geometries. Alternatively, it is possible to operate in a semi-structured environment, defined by the use of softer fixtures (e.g. pickup bin) and softer robotic tools (e.g. suction cups or compliant pads) that can be used for many assembly applications without modification, but they demand specific motion plans that can tolerate greater positional uncertainty. We have developed a system that supports autonomous generation of parameterized fine motion plans for assembly that are robust under positional uncertainty and compliance introduced by the use of a suction cup instead of a gripper. To accomplish this a classifier is trained, implemented and tested for performance in the semi-structured environment for distinguishing between a failed or successful assembly. The trained classifier is then integrated with the entire system and many robot-attended experiments are performed that vary the fine motion parameters, and optimize them for successful outcomes using an Interval Estimation optimization algorithm. An approach to machine learning based on Support Vector Machines and Principal Component Analysis is used to make the optimization autonomous. We achieved a 99.7% classification accuracy with the trained classifier and by running repeated robot-attended experiments with artificial positional uncertainty and optimizing fine motion parameters, we were able to achieve a 38% improvement compared to fine motion plans with initial best guess parameters.