175 0

Balancing control of Bipedal robot using Deep Reinforcement learning

Title
Balancing control of Bipedal robot using Deep Reinforcement learning
Author
Kibeom Kim
Alternative Author(s)
김기범
Advisor(s)
박종현
Issue Date
2019-02
Publisher
한양대학교
Degree
Master
Abstract
The objective of this paper is to apply a Deep Reinforcement Learning(DRL) algorithm, the Deep Deterministic Policy Gradient(DDPG), for a 2D bipedal robot tasks with standing and balancing. For these tasks, this research proposes that new exploration method is derived agent to get stable posture during learning process. The new method is a selective of continuous exploration method for bipedal robot. The traditional exploration method has a limitations which the posture of robot is derived to unstability during learning process. Especially, chattering of joints which is caused by discrete exploration are critical factor to stability of bipedal robot. To overcome these limitations of traditional exploration method, the proposed method which has a feature of combination of continuous exploration and traditional method is developed. The DDPG algorithm is applied on a simulated model of bipedal robot and performance of DDPG is compared new exploration method with traditional exploration method. Reward function which is suitable for bipedal robot is also designed in this paper. The reward function gives positive score to the model when model keeps stable posture sequentially. This method guarantees the stable exploration of robot model in learning process. After finishing learning process, this paper also shows that a bipedal robot which has sensitive feature to get stable posture achieves stability and resists an unexpected disturbance by using DDPG. Consequently, this paper shows better exploration results by the proposed method. To prove this method, the bipedal robot used for this application is DAWRIN-OP which is developed by the Robotis in collaboration with Virginia Tech, Purdue University, and University of Pennsylvania. This approach is simulated for two experimental setups in 2D-plane simulation environment. V-rep which is dynamic simulator, and tensorflow in Python library are used in this simulation.
URI
https://repository.hanyang.ac.kr/handle/20.500.11754/99475http://hanyang.dcollection.net/common/orgView/200000434387
Appears in Collections:
GRADUATE SCHOOL[S](대학원) > MECHANICAL CONVERGENCE ENGINEERING(융합기계공학과) > Theses (Master)
Files in This Item:
There are no files associated with this item.
Export
RIS (EndNote)
XLS (Excel)
XML


qrcode

Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.

BROWSE