Behaviour Aesthetics of Reinforcement Learning in a Robotic Art Installation

Sofian Audry (1) , Rosalie Dumont-Gagné , Hugo Scurto (2)

1 UQAM - Université du Québec à Montréal = University of Québec in Montréal
2 EnsAD - École nationale supérieure des Arts Décoratifs

This paper presents our practice with reinforcement learning (RL) in the context of a robotic art project. We adopted a design-oriented approach to RL, privileging embodied robot behaviours generated by interactive learning processes over offline evaluations of optimal agent policies. We describe aesthetic and technical issues surrounding our experiments with a spheroid robot, and share our reflections on the potential of using RL for behaviour aesthetics.

Domaines: Art et histoire de l'art; Apprentissage [cs.LG]; Robotique [cs.RO]

Sofian Audry, Rosalie Dumont-Gagné, Hugo Scurto. Behaviour Aesthetics of Reinforcement Learning in a Robotic Art Installation. 4th NeurIPS Workshop on Machine Learning for Creativity and Design, Dec 2020, Vancouver, Canada. ⟨hal-03100907⟩