사업진행
[3차연도 정보통신대학원 세미나5] Towards Natural Motion Generator
작성자
고 라경
작성일
2022-05-20 11:36
조회
319
[3차연도 정보통신대학원 세미나5]
▶강사: 최성준 교수
▶제목: Towards Natural Motion Generator
▶초록
Motion generation in Robotics has been widely studied owing to its various applicabilities. For example, when given an explicit object function (e.g., minimizing the energy consumption or finding the shortest path), optimization-based or randomized planning methods are well established for this purpose. However, when it comes to generating natural motions, a relatively small number of studies have been made due to its hardness in defining the 'naturalness' of motion. Throughout this talk, I will cover how AI technologies can be utilized to generate natural motions of both robots (e.g., legged robots and manipulators) and animated characters.
Recently, autonomous robots have been gradually permeating our daily lives, where the necessity of generating natural robotic motions has gained more interest accordingly. However, generating an abundant number of motions for human-like robots from scratch may take an excessive amount of time. Hence, we will first look at how motion capture data can be leveraged to generate humanoid motions, often referred to as motion retargeting. I will also present a robust motion retargeting method to handle noisy motion data estimated from RGB videos (e.g., YouTube). Then, a data-driven motion generation method for animated characters will be presented that can not only generate a natural motion but also stylize the motion. Finally, I would like to share our recent results on perceptive manipulation using reinforcement learning, emphasizing shared autonomy.
▶강사: 최성준 교수
▶제목: Towards Natural Motion Generator
▶초록
Motion generation in Robotics has been widely studied owing to its various applicabilities. For example, when given an explicit object function (e.g., minimizing the energy consumption or finding the shortest path), optimization-based or randomized planning methods are well established for this purpose. However, when it comes to generating natural motions, a relatively small number of studies have been made due to its hardness in defining the 'naturalness' of motion. Throughout this talk, I will cover how AI technologies can be utilized to generate natural motions of both robots (e.g., legged robots and manipulators) and animated characters.
Recently, autonomous robots have been gradually permeating our daily lives, where the necessity of generating natural robotic motions has gained more interest accordingly. However, generating an abundant number of motions for human-like robots from scratch may take an excessive amount of time. Hence, we will first look at how motion capture data can be leveraged to generate humanoid motions, often referred to as motion retargeting. I will also present a robust motion retargeting method to handle noisy motion data estimated from RGB videos (e.g., YouTube). Then, a data-driven motion generation method for animated characters will be presented that can not only generate a natural motion but also stylize the motion. Finally, I would like to share our recent results on perceptive manipulation using reinforcement learning, emphasizing shared autonomy.