Evaluating Deep Reinforcement Learning Algorithms for Physics-Based Musculoskeletal Transfemoral Model with a Prosthetic Leg Performing Ground-Level Walking

Surana, Shikha (2021) Evaluating Deep Reinforcement Learning Algorithms for Physics-Based Musculoskeletal Transfemoral Model with a Prosthetic Leg Performing Ground-Level Walking. Bachelor's Thesis, Artificial Intelligence.

Preview	Text Final Draft - Shikha Surana.pdf Download (1MB) | Preview
	Text toestemming.pdf Restricted to Registered users only Download (104kB)

Abstract

This paper compares deep reinforcement learning algorithms for a physicsbased musculoskeletal osseointegrated transfemoral model with a prosthetic leg performing ground-level walking. The algorithms compared are: Proximal Policy Optimization (PPO) with Imitation Learning, and PPO with Covariance Matrix Adaptation (CMA) and Imitation Learning. The imitation dataset is a public dataset consisting of the joint kinematics of a healthy subject walking in a straight line on flat terrain. Unfortunately, both DRL algorithms were unsuccessful in generating a healthy gait, however, this paper does evaluate them based on: cumulative reward, similarity of kinematics to imitation data, and muscle/actuator usage. Compared to PPO+IL, PPO-CMA+IL received a 78.5% larger mean cumulative reward, 64.7% larger mean duration of an episode, and a decrease of 12.7% in muscle and actuator usage. In contrast, both algorithms performed similar in terms of the proximity between model’s joint kinematics and the imitation data. Future research will refine the reward function and reduce the erraticness of actuator forces.

Item Type:	Thesis (Bachelor's Thesis)
Supervisor name:	Carloni, R. and Raveendranathan, V.
Degree programme:	Artificial Intelligence
Thesis type:	Bachelor's Thesis
Language:	English
Date Deposited:	02 Sep 2021 11:34
Last Modified:	10 Jan 2024 13:27
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/26011

Actions (login required)

View Item