f-Divergence Optimization for Task-Parameterized Learning from Demonstrations Algorithm

Programming robots through demonstration in unstructured environments is often a challenging task, requiring consideration of various parameters. One of the main challenges in an unstructured environment is the ability to extrapolate from user-provided demonstrations, which are sometimes limited. To address this issue, the idea of using task parameterization emerges, assuming that trajectory movements are modulated by different parameters (such as orientation or position) of relevant points, such as the initial and final points of a trajectory. However, some of these task parameters (TPs) may not be relevant for task resolution, especially in environments where various types of movements can occur, introducing additional difficulties in task learning. Additionally, it may happen that two demonstrations contain similar information for task execution (redundancies). This article proposes an approach based on a Task Parameterized Gaussian Mixture Model (TPGMM) for Learning from Demonstrations (LfD) that, through the use of an f-Divergence method (Kullback-Leibler), eliminates redundancy and irrelevance in certain tasks. This allows for a optimal learning model that avoids unnecessary information. The efficiency of the proposed approach has been tested in simulation environments and compared against state-of-the-art algorithms within the LfD domain, demonstrating high efficiency in both cases.