Natural Language Understanding (NLU) is the process of constructing a machine understandable meaning representation from natural language inputs. It is well known that full NLU based on a formal symbolic system in unconstrained environments is an AI-complete problem (Feigenbaum and Feldman, 1995) because no formal symbolic system developed so far, or in the foreseeable future can adequately model the vast background knowledge which humans take for granted in interpreting and reasoning. The focus of most of the current research in NLU is on solving specific problems in narrowly defined domains or micro-worlds . In particular, researchers such as Terry Winograd have successfully produced systems which perform limited tasks based on natural language instructions. However, the most significant limitation of this type of research is that the results produced normally do not generalise well or are not very robust on real life data. Because of this limitation, more research need to be done in the computational modelling of the instructive aspect of natural language on a non-trivial scale. Furthermore, NLU systems based on statistical methods have achieved considerable success in producing sufficiently robust and generalisable results. Therefore, I propose to investigate how to use a formal symbolic system incorporating statistical methods to interpret well-formed English narration which describes physical actions within a preconstructed, totally accessible virtual environment. This research task can be broken down into the following general problems: 1. how to represent the meaning of most of the semantically decomposable verbs using just a small (therefore manageable) group of semantically non-decomposable verbs; 2. how to reliably disambiguate open class words, especially verbs on a large scale in a totally accessible virtual environment; 3. how to reliably capture the formal semantics of natural language inputs; 4. how to utilise the knowledge of the environment to reliably disambiguate ambiguous natural language inputs; 5. how to reliably ground the entities associated with natural language inputs in the environment; 6. how to reliably ground and execute the events described by natural language inputs. While I cannot hope to solve these problems in general, I believe that by carefully constraining the virtual world and natural language inputs, it is possible to make substantial progress in developing a framework and the tools necessary for building specific systems that could be controlled by natural language instructions. I believe that my research has a wide range of applications. It can be adapted individually to control robotic devices in environments whose domain knowledge is sufficiently accessible. It can also be adapted to perform human-computer interaction in restricted domains such as database querying, virtual tour guides and virtual assistants. In my research, I will focus on the application of controlling virtual characters on a pre-constructed virtual stage.