A computer learning system is presented, which is capable of gaining and augmenting the operational skills for a mobile robot to navigate in unexplored environments. Rather than having the robot programmed to work in a designated environment, it is trained to acquire generic knowledge about visual guidance and motion control in the environments which bear some common features of the learning environment. In the training process, the robot learns to make decisions under different circumstances, represented as meta-scenes. The learning process applies an interactive instruction method and a generic rule modification scheme. An explanation-based reasoning approach and a pattern-matching technique for knowledge induction are implemented. The learning system structure and computer simulations are illustrated.