| On 2023-11-21 at 15:00:00 (Brussels Time) |
Abstract
The analysis of robot swarms aims to ensure that their behaviour holds some desired properties. Generally, a macroscopic model of the swarm is used to provide analytically tractable guarantees. The challenge of building suitable models, however, limits their applicability. I will present a data-driven representation of robot swarm behaviours that lends itself to mathematical analysis. I will show how to represent behaviours of spatial organisation and navigation as mesoscopic vector fields. I will then show how the natural Helmholtz-Hodge decomposition can be used to extract necessary topological features of the behaviour, as well as to quantify and assess its degradation when the corresponding control software is executed in a different simulated environment. I will showcase this approach on a set of control software instances designed with AutoMoDe-Chocolate and EvoStick, two automatic design methods. The results show that EvoStick produces control software that is more prone to topological degradation, where the features of the behaviour change.
