Arthroscopic surgery has become increasingly common in recent years. In this surgical technique, the visual field is extremely limited and the degrees of freedom in operation are lower than that in open surgery. An immature surgical technique could run the risk of damaging articular cartilage and causing other complications, and there only a few reported systems for the objective evaluation of surgical skills in arthroscopic surgery training programs. We propose a training system for arthroscopic surgery using a newly developed box trainer, which is believed to provide more efficient training for surgeons. The box trainer has ten kinds of training exercises. Each task is aimed at training specific skills required for arthroscopic surgery, using simple geometrical shapes or realistic anatomical shapes. The box trainer is controlled by a microcomputer and is able to inform trainees of score, working time and the number of errors. Trainees are notified with a beep alarm when errors are performed. Knee models are fabricated for these tasks using anatomical information from a magnetic resonance imaging knee data set and a 3D reconstruction algorithm with computer-aided manufacturing .