Behaviour is the ultimate output of an animal's nervous system and choosing the right action at the right time can be critical for survival. The study of the organisation of behaviour in its natural context, ethology, has historically been a primarily qualitative science. A quantitative theory of behaviour would advance research in neuroscience as well as ecology and evolution. However, animal posture typically has many degrees of freedom and behavioural dynamics vary on timescales ranging from milliseconds to years, presenting both technical and conceptual challenges. Here we review 1) advances in imaging and computer vision that are making it possible to capture increasingly complete records of animal motion and 2) new approaches to understanding the resulting behavioural data sets. With the right analytical approaches, these data are allowing researchers to revisit longstanding questions about the structure and organisation of animal behaviour and to put unifying principles on a quantitative footing. Contributions from both experimentalists and theorists are leading to the emergence of a physics of behaviour and the prospect of discovering laws and developing theories with broad applicability. We believe that there now exists an opportunity to develop theories of behaviour which can be tested using these data sets leading to a deeper understanding of how and why animals behave.