Performance-based Adaptation of Lateral Robotic Balance Assistance during Slackline Walking

Master thesis (2020)

Authors

A. Aneesh Ashok Kumar Mechanical Engineering

Contributors

Andrew Berry Biomechatronics & Human-Machine Control - Mechanical, Maritime and Materials Engineering (mentor)
H. Vallery Biomechatronics & Human-Machine Control - Mechanical, Maritime and Materials Engineering (mentor)
J.C.F. de Winter Human-Robot Interaction - Mechanical, Maritime and Materials Engineering (coach)
Faculty
Mechanical Engineering, Mechanical Engineering
Copyright: © 2020 Aneesh Aneesh Ashok Kumar
More Info
expand_more

Published Date

14-05-2020

Language

English

Reuse Rights

Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.

Faculty

Mechanical Engineering

Abstract

As part of the ongoing research on balance assistance at the Delft Biorobotics Lab, the goal of this project was to analyse the relevance of optimising the control action to each individual user. In order to explore the feasibility of tuning balance assistance based on task performance as well as controller assistance, an experimental study was conducted on the optimisation of lateral balance assistance during slackline walking. 
The task performance was quantified as the distance walked on the slackline. The balance assistance was provided by the RYSEN– an active Bodyweight Support (BWS) system used in gait training– in the form of lateral damping. The damping value was tuned over successive trials to optimise the trade-off between maximising the distance walked by the participants before loss of balance and minimising the lateral impulses applied by the RYSEN during each trial. The Covariance Matrix Adaptation Evolution Strategy (CMAES) was used for the
automated optimisation of the lateral damping assistance. The performance of the CMAES algorithm was compared to participant performance in optimising the damping based on the same cost measure.
Using a final sample size of 15 healthy participants, comparisons were made between costs for trials conducted with Zero Damping (ZD), the optimal damping estimate returned by the algorithm (Da) and the optimal damping estimate selected by the participants (Dh), with the expectation that lower trial costs would be returned by the optimal damping estimates.
No significant difference was found among the median cost measures of the three conditions (p = 0.24, Friedman’s ANOVA). Significant differences were found between the corresponding distance measures (p = 0.002, Friedman’s ANOVA), with significantly longer distances walked in Da and Dh trials compared to ZD trials (p = 0.003 and p = 0.005 respectively), but no significant differences between the measures for Da and Dh trials (p = 0.61). Analysis of the primary cost measures based on Grid Search (GS) trials indicated a shallow cost landscape with high variability. 
Post-hoc optimisation simulations were run using CMAES and Bayesian optimisation upon individual participant grid search data. Results indicated the possibility of better performance upon increasing the number of candidates sampled per generation and the initial step size used in the algorithm, using an alternative cost function with a quadratic distance cost component, or by using Bayesian optimisation for a comparable number of trial iterations.