Generalizing kinematic skill learning to energy efficient dynamic motion planning using optimized Dynamic Movement Primitives 

Tian Xu*,  Siddharth Singh*, and Cindy Chang

*Equal Contribution

Robotics and Computer Integrated Manufacturing, Elsevier

Abstract

In manufacturing, automating the generation of dynamic trajectories for di verse robots and loads in response to kinematic task requirements presents a significant challenge. Previous research has primarily addressed kine matic trajectory generation and dynamic motion planning as separate en deavors, with integrated solutions rarely explored. This paper presents a novel methodology that combines reinforcement learning (RL)-based kine matic skill learning, dynamic modeling and an enhanced version of Dynamic Movement Primitives (DMP). Utilizing a pre-established skill library, the RL-enabled method generates multiple kinematic trajectories that fulfill the specific task requirements. These trajectories are further refined by dynamic modeling, selecting paths that minimize energy consumption tailored to spe cific robot types and loads. The newly proposed Optimized Normalized Dynamic Motion Primitive (ON-DMP) enhances obstacle avoidance with min imal energy consumption, remaining effective in novel environments. Vali dated through both simulated and real-world experiments, this methodology shows robust results in improving task completion in dynamic real-world environments without the need of reprogramming .

Graphical Abstract

Experiment Video

@article{XU2025102983,

title = {Generalizing kinematic skill learning to energy efficient dynamic motion planning using optimized Dynamic Movement Primitives},

journal = {Robotics and Computer-Integrated Manufacturing},

volume = {94},

pages = {102983},

year = {2025},

issn = {0736-5845},

doi = {https://doi.org/10.1016/j.rcim.2025.102983},

url = {https://www.sciencedirect.com/science/article/pii/S0736584525000377},

author = {Tian Xu and Siddharth Singh and Qing Chang},

keywords = {Motion planning, Dynamic movement primitives, Learning from demonstration}

}