Force feedback plays an important role in collaborative virtual reality environments, mainly for programmers of haptic visualization tools. Whereas a great deal of work has gone into graphical displays over the past years, little has changed on the input side. One of the problems that has slowed down development in this area is the difficulty of integrating the visualization of a scene, the interaction of the user with the scene, the feeling for the user to be immersed inside the scene, and finally, the input devices. In this paper, we describe the architecture we have designed, implemented and tested for a collaborative virtual training using force feedback devices. In particular, it provides device independence and easy extensibility through a compartmentalized and multilayered model. We also present examples of how force feedback joysticks can be integrated into training exercises using our prototype.