Computational neuroscience has become increasingly reliant on high-performance computing as the scale and complexity of models have grown. The increased complexity also increases the need to test models with real-life inputs and to interpret model outputs in functional and behavioural terms. This can be achieved by embedding the models in the real world using robotic hardware. We present an experimental test of an embedded, large-scale simulation, where a robot was remotely connected to a functional network model of the early saccade generation circuit, running in a cluster, completing a perception-action loop. We discuss the technical and organisational issues that need to be addressed for this kind of embedded high-performance modelling to become feasible.