Cultivating Quadrupedal Robotic Agility: Unveiling Reinforcement Learning Dynamics with the Boosted Tree Motif Classifier

This study examines the Boosted Tree Motif Classifier (BTMC), which leverages gradient boosting for the precise classification of intricate motion patterns in quadrupedal robots, with a particular emphasis on dynamic actions like jumping. Building upon our previous conference paper [C. Allred, J. Pusey and M. Harper, Detecting ballistic motions in quadruped robots: A BTMC for understanding reinforcement learning, in 7th IEEE Int. Conf. Robotic Computing (IEEE, 2023)], this journal extension expands the study by incorporating multivariate motif analysis, additional machine learning models, and a more comprehensive evaluation of reinforcement learning dynamics. BTMC achieves 96% precision, effectively identifying complex motion patterns during training. Through experiments at multiple levels of from individual motors to full robotic actuation, BTMC consistently outperforms traditional techniques in accuracy. This study lays the foundation for future research on whole-body reward development, providing clearer insights into the intricate relationship between rewards and learned skills.