Cited time in webofscience Cited time in scopus

Full metadata record

DC Field Value Language
dc.contributor.advisor 임용섭 -
dc.contributor.author Yaeohn Kim -
dc.date.accessioned 2023-03-22T19:56:42Z -
dc.date.available 2023-03-23T06:00:22Z -
dc.date.issued 2023 -
dc.identifier.uri http://hdl.handle.net/20.500.11750/45710 -
dc.identifier.uri http://dgist.dcollection.net/common/orgView/200000652227 -
dc.description Dynamic window approach, Deep reinforcement learning, Collision avoidance, Local path planner, Autonomous mobile robot -
dc.description.abstract Path planning for autonomous mobile robot is considered as the core algorithm to avoid static and dynamic obstacles. Thus, many recent researches were achieved with dynamic window approach (DWA) as local path planner, whose combination of weighting parameters could easily generate different path planning strategies. However, the performance of DWA highly depends on such combination, and requires large amount of time to find the optimal value through trial and error. To solve these drawbacks, we implemented deep reinforcement learning (DRL) network model based on DWA to dynamically adapt parameters. By the proposed approach, we took the advantages of time efficiency and secured safety compared to full replacement of path planning algorithm as end-to-end learning algorithm. Furthermore, we proposed a new tunable parameter in DRL based DWA and trained our algorithm based on our network model. As a result, collision free action of the mobile robot was strongly guaranteed, and relatively high success rate was successfully observed in simulations. Moreover, we have tested our path planning model on real world environment and showed improved performance in obstacle avoidance than other DWA algorithms that had set tunable parameters as a constant value. -
dc.description.abstract 본 논문은 심층 강화 학습 기법을 지역 경로 계획 알고리즘 중 하나인 동적 창 접근 방식에 적용하여 실시간 장애물 회피 주행을 하는 알고리즘을 제공한다. 기존 동적 창 접근 방식을 수정하여 회피 주행 능력을 결정하는 기존 가중치 변수 외에 또 다른 하나의 변수를 제안한다. 강화 학습의 관찰 영역은 모바일 로봇의 선속도와 각속도이며, 행동 영역은 기존 가중치 변수 3개와 제안된 변수 하나이다. 제안한 강화 학습 모델과 보상 함수는 2개의 층으로 구성된 신경망을 통해 차륜 구동 모델의 모바일 로봇과 장애물이 포함된 랜덤한 시나리오가 구현된 가상환경 속에서 학습을 거친다. 가장 잘 학습된 모델을 시뮬레이션 평가 환경과 실험 환경에 적용한다. 본 연구에서 제안한 실시간 장애물 회피 주행 알고리즘은 기존 동적 창 접근 방식과 같이 성공적으로 회피 주행하는 것을 관찰할 수 있다. -
dc.description.tableofcontents Ⅰ. Introduction 1
Ⅱ. Related Works 3
Ⅲ. Problem Statement 5
3.1 Dynamic Window Approach 5
3.2 Adaptive Prediction Length based Dynamic Window Approach 8
3.3 Deep Reinforcement Learning Policy Training 12
Ⅳ. Comparison and Evaluation 15
4.1 Simulation Results 15
4.2 Experimental Results 18
Ⅴ. Conclusions, Limitations and Future Works 20
References 21
국문 초록 23
-
dc.format.extent 23 -
dc.language eng -
dc.publisher DGIST -
dc.title Collision Avoidance for Autonomous Mobile Robot: Adaptive Prediction Length based Dynamic Window Approach with Deep Reinforcement Learning -
dc.title.alternative 심층 강화 학습을 적용한 동적 창 접근 방식 기반의 장애물 회피 알고리즘 개발 -
dc.type Thesis -
dc.identifier.doi 10.22677/THESIS.200000652227 -
dc.description.degree Master -
dc.contributor.department Department of Robotics and Mechatronics Engineering -
dc.contributor.coadvisor Sukho Park -
dc.date.awarded 2023-02-01 -
dc.publisher.location Daegu -
dc.description.database dCollection -
dc.citation XT.RM 김64 202302 -
dc.date.accepted 2023-03-21 -
dc.contributor.alternativeDepartment 로봇및기계전자공학과 -
dc.subject.keyword Dynamic window approach -
dc.subject.keyword Deep reinforcement learning -
dc.subject.keyword Collision avoidance -
dc.subject.keyword Local path planner -
dc.subject.keyword Autonomous mobile robot -
dc.contributor.affiliatedAuthor Yaeohn Kim -
dc.contributor.affiliatedAuthor Yongseob Lim -
dc.contributor.affiliatedAuthor Sukho Park -
dc.contributor.alternativeName 김예온 -
dc.contributor.alternativeName Yongseob Lim -
dc.contributor.alternativeName 박석호 -
Files in This Item:

There are no files associated with this item.

Appears in Collections:
Department of Robotics and Mechatronics Engineering Theses Master

qrcode

  • twitter
  • facebook
  • mendeley

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

BROWSE