Improved Gaussian mixture model and Gaussian mixture regression for learning from demonstration based on Gaussian noise scattering

Chunhua Feng, Zhuang Liu, Weidong Li, Xin Lu, Yanguo Jing, Yongsheng Ma

Abstract: Learning from Demonstration (LfD) is an effectual approach for robots to acquire new skills by implementing intuitive learning through imitating human demonstration. As one of the mainstream learning models for LfD, Gaussian mixture modeling (GMM) and Gaussian mixture regression (GMR) exhibit the advantages of ease of use and robust learning capabilities. To further improve the learning and regression performance of GMM/GMR, in this paper, improved GMM/GMR based on a Gaussian noise scattering strategy is designed. The main contributions of this study include: 1) the Gaussian noise scattering strategy is developed to eliminate the requirement of creating multiple demonstrations and overcome the jitter and sharp-turning defects of the demonstration; 2) based on a new evaluation criterion IBF and the sparrow search algorithm (SSA), GMM/GMR is optimized to achieve the balance of feature retention of the demonstration and the smoothness of the reproduced solution. Experimental results show that with the Gaussian noise scattering strategy, the geometric similarity of the reproduced solution and the demonstration increased for approximately 33.16 %, and the smoothness improved for 19.83 %. The challenges of underfitting and overfitting in GMM/GMR were effectively mitigated after incorporating the evaluation criterion IBF and leveraging SSA. This demonstrates the potential applicability of the improved GMM/GMR in practical industrial scenarios.

External IDs:dblp:journals/aei/FengLLLJM25