Quantifying Generalisation in Imitation Learning