Crowd simulation is an important cornerstone for effective urban planning and emergency management. Accuracy is crucial, and good empirical data is required to build and calibrate accurate simulation models. Unfortunately, it is very difficult to obtain good data on crowd movements, particularly under emergency and panic conditions. Realistic experiments under panic conditions are impossible, chance recordings are rare and incomplete, and experiments under standard conditions are likely to be only marginally relevant. In this project we will be using physics-based methods to investigate which types of models are adequate abstractions that can drive useful simulations.