Welding process parameters are indispensable to program arc welding robot. To simplify off-line programming (OLP) for robotic arc welding, we develop an arc welding expert system whcih can generate welding process parameters automatically. Its input data came from the feature database of welding part, which is set up by our feature modeling system. The expert system has become an important module of our RAWTOLPS (Robotic Arc Welding Task-level Off-Line System). It combines case-based reasoning with heuristic rule-based reasoning methods to deal with the welding process design. Moreover, artificial neural networks are introduced to the systems for reasoning and machine learning, and several network modules are developed to learn from welding process database, based on back-propagation neural networks. After some groups of actual welding process data were used to train the network models, several network models are established to both design the welding process and to predict the weld bead shape. Besides the ANN-based learning, cased-based learning are used in the expert system. These two methods have respectively their own characteristics, and can meet qualifications of different users. The experimental data show that the system can accomplish re-learning and expanding of welding process knowledge, and satisfy the command of the off-line programming system.