University of Twente Student Theses

Login

Insights on human standing balance based on deep learning-driven musculoskeletal simulations

Trigo La Blanca, Carlota (2023) Insights on human standing balance based on deep learning-driven musculoskeletal simulations.

[img] PDF
14MB
Abstract:During the first year of human life, the ability to stand and maintain balance is acquired effortlessly, yet there are still gaps in the understanding of how humans perform those tasks. This thesis aims to enhance the comprehension of human movement control by training a reinforcement learning policy that controls a musculoskeletal model of the lower limb to perform standing balance with and without perturbation tasks. A reward function was designed based on the pelvis position error, the stability margin, and the metabolic cost. The observation space was composed of joint position, velocity, pelvis position, pelvis error, and muscle activations. Training for standing balance entailed 15 million steps. For the standing balance with perturbation task, a randomized force between 1 and 50 N was applied to the pelvis for 0.1 seconds. Two separate 20-million-step training were conducted, with the perturbation applied in the anterioposterior and mediolateral directions, respectively. The results align with existing literature, although the lower limb model does not replicate the expected human response. The model withstands higher anterioposterior perturbations compared to mediolateral perturbations, using the ankle strategy for both types of perturbations. Overall, the myoLeg model is able to stand and withstand perturbations using solely information on joint position, velocity, pelvis position, error, and muscle activation information.
Item Type:Essay (Master)
Faculty:ET: Engineering Technology
Subject:01 general works
Programme:Biomedical Engineering MSc (66226)
Link to this item:https://purl.utwente.nl/essays/97268
Export this item as:BibTeX
EndNote
HTML Citation
Reference Manager

 

Repository Staff Only: item control page