Novelty occurs all the time in the world: new things appear, and known things change. Robust AI agents acting in an open world must be able to recognise novelty, characterise it, and adapt to it.

We explore novelty in the context of the physics-simulation puzzle game Angry Birds. Despite its simple appearance, solving this game excercises many aspects of the same reasoning capabilities that will be needed by AI agents acting in the real world, and it is the platform for an AI challenge and competition. The images below show an example of how novelty can appear: In the normal game (top row), the red bird falls to the ground after hitting (and destroying) the stone block. In the right image, the behaviour of the bird has changed, and it instead bounces off the block.

The goal of this project is to develop a system that computes a characterisation of novelty, in the form of a description in a formal knowledge representation language. From observation of the standard game being played, we extract a description of what can normally happen. This description uses a qualitative abstraction of spatial relations in the world, i.e., concepts like "left of", "moving towards", "touching", rather than concrete positions and velocities. Novelty becomes apparent when this description no longer matches what is observed, and the characterisation is then made by computing the changes to the description that are needed to cover the new observations.