Overcoming some drawbacks of Dynamic Movement Primitives

Dynamic Movement Primitives (DMPs) is a framework for learning a point-to-point trajectory from a demonstration. Despite being widely used, DMPs still present some shortcomings that may limit their usage in real robotic applications. Firstly, at the state of the art, mainly Gaussian basis functions have been used to perform function approximation. Secondly, the adaptation of the trajectory generated by the DMP heavily depends on the choice of hyperparameters and the new desired goal position. Lastly, DMPs are a framework for 'one-shot learning', meaning that they are constrained to learn from a unique demonstration. In this work, we present and motivate a new set of basis functions to be used in the learning process, showing their ability to accurately approximate functions while having both analytical and numerical advantages w.r.t. Gaussian basis functions. Then, we show how to use the invariance of DMPs w.r.t. affine transformations to make the generalization of the trajectory robust against both the choice of hyperparameters and new goal position, performing both synthetic tests and experiments with real robots to show this increased robustness. Finally, we propose an algorithm to extract a common behavior from multiple observations, validating it both on a synthetic dataset and on a dataset obtained by performing a task on a real robot. (C) 2021 Elsevier B.V. All rights reserved.

Automated summary

Dynamic Movement Primitives (DMPs) framework is widely used for learning point-to-point trajectories, but still has shortcomings in terms of function approximation methods, trajectory adjustment dependencies, and constraints on one-shot learning.