Regensburg 2022 – wissenschaftliches Programm
Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
CPP: Fachverband Chemische Physik und Polymerphysik
CPP 22: Active Matter 3 (joint session BP/CPP/DY)
CPP 22.1: Vortrag
Mittwoch, 7. September 2022, 09:30–09:45, H16
Collective foraging of microrobots trained by reinforcement learning — •Robert C. Löffler1, Emanuele Panizon2, and Clemens Bechinger1 — 1Fachbereich Physik, Universität Konstanz, Konstanz, Germany — 2Department of Quantitative Life Science, International Centre for Theoretical Physics, Trieste, Italy
From bacteria to mammals, collective behavior can be observed on all scales in nature. It is generally driven by the benefit to individuals when cooperating with others. However, the exact motivation of individuals to participate is challenging to investigate, as biological creatures are complex systems theirself. At the same time engineers seek to create collective groups of autonomous systems to perform dedicated tasks by cooperation.
Here we present an experimental model system of feedback-controlled microswimmers which are trained with multi agent reinforcement learning in an actor-critic scheme. A group of active particles is situated in a 2D environment containing a virtual food source which is changing position over time. Despite being rewarded individually for being inside the food source, particles show cohesive collective motion forming flocks and swirls. This is driven by the benefit of social information and collision avoidance, resulting in faster migration to a relocated food source. Understanding those mechanisms behind the emergence of collective behavior is of biological interest as well as to understand human crowd behavior and to design future robotic systems.