Work

Learning rhythmic movements by demonstration using nonlinear oscillators

Year: 2003
Type: conference-paper
Abstract: This paper presents a new approach to the generation of rhythmic movement patterns with nonlinear dynamical systems.Starting from a canonical limit cycle oscillator with well-defined stability properties, we modify the attractor landscape of the canonical system by means of statistical learning methods to embed arbitrary smooth target patterns, however, without losing the stability properties of the canonical system.In contrast to non-autonomous movement representations like splines, the learned pattern generators remain autonomous dynamical systems which robustly cope with external perturbations that disrupt the time flow of the original pattern, and which can also be modified on-line by additional perceptual variables.A simple extension allows to cope with multiple degrees-of-freedom (DOF) patterns, where all DOFs share the same fundamental frequency but, otherwise, can move in arbitrary phase and amplitude offsets to each other.We evaluate our methods in learning from demonstration with an actual 30 DOF humanoid robot.Figure -8 and drumming movements are demonstrated by a human, recorded in joint angle space with an exoskeleton, and embedded in multi-dimensional rhythmic pattern generators.The learned patterns can be used by the robot in various workspace locations and from arbitrary initial conditions.Spatial and temporal invariance of the pattern generators allow easy amplitude and speed scaling without losing the qualitative signature of a movement.This novel way of creating rhythmic patterns could tremendously facilitate rhythmic movement generation, in particular in locomotion of robots and neural prosthetics in clinical applications.
Language: English
Open Access status: green
OpenAlex ID: W2115817551
FWCI: 23.76
Cites: 14
Cited by: 188
Related to: 10