In swarm robotics each robot exchanges its control parameters, which quality is assessed by an on-board objective function, and the objective is to display microscopic behaviour that combined together give rise to a desired macroscopic organisation. The learning abilities of the swarm of robots is related to the dynamics of their internal parameters in an unknown and complex energy landscape. 

On the other hand, from the point of view of physics, a swarm of robots can be seen as a fluid, composed of self-propelling units, a so-called “active liquid”, in which the individual units carry a discrete set of interacting policy's parameters. 

The goal of this thesis is to investigate how distributed online learning process mimicking social learning, can be used to discover and exploit primitives available at the physical level. the broader objective is to propose an interdisciplinary viewpoint of swarm robotics, in order to leverage what can be freely provided by physics, when designing computational decision making algorithm. 

We propose to investigate numerically a behavioural learning process, distributed on a set of mobile agents, as one would do for an active liquid of particles carrying internal degrees of freedom, which couple dynamically to the motion of the particles. We will investigate the interplay between the learning process, which evolves the dynamical rules of the agents, and the underlying physical primitives, which condition the large scale phases available to the swarm. The expected outcome is both a practical learning algorithm for dense swarm robotics, and theoretical results about the dynamics of collective learning in a physical setup. 

 

PhD student: Jérémy Fersulat 

PhD supervisor: Nicolas Bredeche

Research laboratory: ISIR - Institut des Systèmes Intelligents et de Robotique