Overcoming some drawbacks of Dynamic Movement Primitives

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