CO2 lasers are increasingly being utilized for quality welding in production. Considering the high cost of equipment, the start-up time and the set-up time should be minimized. Ideally the parameters should be set up and optimized more or less automatically. In this paper a control system is designed and built to automatically optimize the focal point position, one of the most important parameters in CO2 laser welding, in order to perform a desired deep/full penetration welding. The control system mainly consists of a multi-axis motion controller - PMAC, a light sensor - Photo Diode, a data acquisition card - DAQCard-700, and a self-learning mechanism - Neural Network. The optimization procedure starts with the welding process being carried out by continuously moving the focal point position from above a welding plate to below the plate, thus the process is ensured to be shifted from initially surface welding to deep/full penetration welding and back to surface welding again. A clear change on plasma brightness from the process is monitored by the photo diode on the front side of the plate with a viewing angle of 45o. The photo diode signal is acquired with the A/D converter card and installed in a computer hard disk for later data processing. Thereafter the optimum focal point position (OFPP) is output by the self-learning mechanism - the neural network. The optimization procedure is completed with the welding process being carried out by adjusting the focus of the laser beam to the OFPP.A self-learning mechanism - neural network as the essence of the control system is trained with the photo diode signals extracted from various welding processes with the changes on the laser power, translation speed, material and thickness of the plate, shielding gas type and flow rate, and welding configuration. The results of the self-learning focus control system show that the neural network is capable of optimizing the focal point position with good accuracy in CW CO2 laser welding.