Sample-Efficient Motion Generalization Via Task-Parameterized GMMs with Bayesian Hyperparameter Optimization

Learning from Demonstration (LfD) is a promising approach to teaching robotic manipulators by example, enabling intuitive and efficient behavior acquisition. However, most existing methods struggle with data efficiency and generalization, particularly in scenarios where only a few demonstrations are available. In this work, we propose an architecture that combines Task-Parameterized Gaussian Mixture Models (TP-GMM) with Bayesian Optimization (BO) for automatic hyperparameter tuning. TP-GMM enables the encoding of demonstrations in multiple local reference frames, capturing the relative dynamics of the task and facilitating generalization to unseen configurations. BO, in turn, automatically selects key model parameters, such as the number of components, covariance regularization, and input-output weights, by maximizing model performance via cross-validation. We validate the proposed approach in a real-world pick-and-place manipulation task using a Kinova Gen3 robot, with varying task parameters inferred through visual markers. By relying on only five demonstrations per subtask, our method achieved a 95% task success rate, outperforming both standard TP-GMM and a TP-GMM+EKF variant in terms of accuracy and consistency. These results demonstrate that our method outperforms non-parameterized models in accuracy and robustness while maintaining low computational cost and high sample efficiency, highlighting its potential for scalable and generalizable robotic learning.