Description

Abstract: Curriculum learning has emerged as a promising approach for improving motor skill
acquisition in reinforcement learning tasks. This study investigates the efficacy of a speed-first
curriculum in a throwing motion control task using a single-link planar pendulum. Drawing on the
resolution of ill-posed problems in motion control and the inverse U-shaped relationship between speed and accuracy in ballistic skill acquisition, the speed-first curriculum was designed to prioritize rapid exploration of the state space. Numerical experiments demonstrate that this approach outperforms accuracy-first and no-curriculum baselines, achieving superior sampling efficiency, generalization capability, and near-zero failure rates in subtasks such as the swing-up motion. The findings underscore the importance of task segmentation and curriculum design in guiding learners toward optimal solutions, particularly in high-dimensional control systems. While this study focuses on low-dimensional systems, future research should explore the extension of these principles to articulated controllers with redundant degrees of freedom, such as those observed in human throwing motions.These results contribute to a broader understanding of curriculum learning and its potential to enhance motor skill acquisition, offering a foundation for applications in robotics and artificial intelligence.

Keywords: Deep Reinforcement Learning Motion Control; Curriculum Learning; Speed and Accuracy