A Unified Coadaptation Framework for Continuous Control

Loading...
Thumbnail Image

URL

Journal Title

Journal ISSN

Volume Title

Sähkötekniikan korkeakoulu | Master's thesis

Date

2023-08-21

Department

Major/Subject

Control, Robotics and Autonomous Systems

Mcode

ELEC3025

Degree programme

AEE - Master’s Programme in Automation and Electrical Engineering (TS2013)

Language

en

Pages

56

Series

Abstract

Coadaptation in robotics studies the joint adaptation of structural and behavioral characteristics. The goal is to develop methods for automating the process of designing a robot. In the current state of research, many coadaptation algorithms exist to optimize a robot's morphology and behavior. The reinforcement learning (RL) framework is well recognized for solving the coadaptation problem. Most RL benchmarking frameworks, such as OpenAI's Gym, only consider the problem of optimizing behavior and therefore do not provide API for optimizing a robot morphology. Hence researchers have to resort to making their benchmarks to validate their coadaptation method. However, this results in various benchmarks and performance metrics, making it hard to compare different coadaptation algorithms. This thesis presents a benchmarking framework for the general use of coadaptation in robotics. This framework provides an API for changing the morphological parameters of the robot whilst keeping the topology fixed. With this framework, two experiments were conducted. Firstly, an investigation into the design space was taken, and a map is produced showing the relationships between morphological parameters and the overall performance of the design. Furthermore, the first benchmark for the framework is presented. In its current state, the framework serves as a springboard for further comparisons between coadaptation methods and other experiments in coadaptation.

Description

Supervisor

Kyrki, Ville

Thesis advisor

Luck, Kevin

Keywords

robotics, evolutionary robotics, coadaptation, co-optimization, reinforcement elarning, benchmarking

Other note

Citation