Traditionally, the focal point in the design of a learning game is the determination of learning objectives and the modelling of scenarios allowing for the objectives to be reached by the learner. In an immersive collaborative multi-character learning game, human players and AI-controlled characters are expected to collaborate and communicate in a shared virtual environment. From the designer’s perspective, the challenge of designing the scenario and having a team of characters follow them as expected is raised. In this paper, we introduce a methodology where the environment is made a central design component of the game. The idea underlying our approach is to model and use a semantic representation of the environment where every object and characteristic playing a role in the game scenarios is represented in an easily accessible and non-ambiguous fashion. This unique representation is capable of guaranteeing at once that the learning scenarios will seamlessly be applied and that players and non playing characters will effectively collaborate with each other within the same 3D environment. Besides, the approach also enables to monitor the activity of each player and deliver a personalised performance assessment against the learning objectives. This paper is presented as a case study where practical examples from 3DVOR, a medical learning game currently under development, are provided to illustrate and to demonstrate the usability of the methodology.