Abstract

The task of accurately throwing a paper plane poses significant challenges in the realm of dynamic robotic manipulation, demanding adaptability to unpredictable aerodynamic properties of paper planes. This paper presents a novel approach to accurate paper plane throwing using reinforcement learning and trajectory encoding. We introduce a method that combines a Variational Au toencoder (VAE) for encoding paper plane trajectories with a Soft Actor-Critic (SAC) algorithm to learn optimal throwing strategies. Our approach dynamically adapts to the unique aerodynamic properties of randomly generated paper plane designs. Our preliminary experiments demonstrate that incorporating information from previous throws improves performance, particularly when generalizing to unseen plane designs. By addressing the complexities of this task, our work has the potential to advance the learning of dynamic robotic manipulations.