Good Vibes: A Reinforcement Learning strategy to optimally actuate vibratory feeders

Vibratory feeders are essential for part presentation in manufacturing but require time-consuming manual tuning. This paper proposes a reinforcement learning strategy that autonomously selects vibration parameters from visual feedback to achieve optimal part singulation. An actor-critic network learns from images describing part dispersion on a flat tray, without prior system modelling. Experiments on a custom setup with an ABB YuMi robot show that the learned policy improves separation efficiency and repeatability compared to a fixed-parameter baseline, enabling adaptive, vision-based feeding for flexible production.