Multi-Objective Bilevel Bayesian optimization for robot and behavior co-design

Traditionally, the robot design process is based on the trial-and-error approach that involves repeated cycles of design, prototyping, and evaluation. During the process, the robot designer should tackle multiple objectives, which are commonly in conflicting relationships. Furthermore, the robot design assessment involves costly behavior optimization and performance evaluation in multiple environments. We propose a Multi-Objective Bilevel Bayesian optimization (MO-BBO) algorithm to automate the co-design process of the robot design and behavior simultaneously. Since the behavior should be optimized for each design and environment, we select the next design candidate in a bilevel manner. Design parameters and behavior parameters are the high- and low-level decision variables, respectively. We applied our algorithm to two robot co-design problems: gripper design problem and robot arm placement problem. MO-BBO was able to effectively expand the Pareto front in objective space on two problems. We extend the robot arm placement problem by integrating human-likeness into objectives. To account for human likeness, we construct trajectory-based metrics to evaluate how well the robot arm follows the human motion trajectories extracted from the TUM Kitchen dataset and how similar the robot arm structure is to the human arm structure while following the trajectories. Compared to the designs generated by using reachability indices, the designs generated by the trajectory-based metrics have better performance when following human motion trajectories, especially in terms of collision rate and structural similarity.

Name thesis:degree_name

M.S.

Level thesis:degree_level

Thesis

Discipline thesis:degree_discipline

Mechanical Engineering

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2022

Author dc:creator

• Kim, Yeonju

Contributors dc:contributor

• Hauser, Kris
• Ramos, João

• Robot Design Optimization
• Robot Arm Placement
• Gripper Design

Statement dc:rights

• Copyright 2021 Yeonju Kim

Language dc:language

en

Handle dc:identifier

http://hdl.handle.net/2142/112968

OAI identifier oai:identifier

oai:www.ideals.illinois.edu:2142/112968