Flock heterogeneity and its applications

Although the vast majority of synthetic flocking models assume that agents in a flock are homogeneous, biologists have shown that this is not the case in real flocks. Mixed-species flocks, for example, are quite common, as are differences in interaction behaviours between agents in a flock. This heterogeneity presents a barrier to our understanding of flocking behaviour, and must be taken into consideration when developing models of such behaviour. This thesis makes the following contributions. First, it describes a software tool for generating photo-realistic images of synthetic flocks for the purpose of training a neural network to learn individual-level attributes such as bird species, position, depth, and flapping phase. Although a quantitative evaluation of this tool is not available at the time of writing, qualitative analysis shows that the output from this tool shares many features in common with photos of real flocks. The second contribution of this thesis is to describe and evaluate the performance of an engineered flocking controller that exploits the pairwise flocking phenomenon---the tendency for birds of certain species to flock in stable monogamous pairs. Advantages of this engineered controller include more stable flock formations that are less likely to break up in the presence of obstacles, as well as a reduced number of tracking interactions overall between agents. This controller is evaluated using both simulated and real drones (on a small team of novel low-cost drones called the uBee). The aforementioned advantages are found to be present in both settings at multiple scales. These results show that an engineered pairwise flocking controller may have useful real-world applications in settings where flock stability or limited computational resources available for tracking flock mates are important factors.