Harlan, Brain (2024) Hybrid multi agent navigation in warehouse environments using CBS-MARL. Masters thesis, Universitas Pelita Harapan.
![Title [thumbnail of Title]](http://repository.uph.edu/style/images/fileicons/text.png)
Title_BRAIN HARLAN.pdf
Available under License Creative Commons Attribution Non-commercial Share Alike.
Download (107kB)
![Abstract [thumbnail of Abstract]](http://repository.uph.edu/style/images/fileicons/text.png)
Abstract_BRAIN HARLAN.pdf
Available under License Creative Commons Attribution Non-commercial Share Alike.
Download (188kB)
![ToC [thumbnail of ToC]](http://repository.uph.edu/style/images/fileicons/text.png)
ToC_BRAIN HARLAN.pdf
Available under License Creative Commons Attribution Non-commercial Share Alike.
Download (208kB)
![Chapter 1 [thumbnail of Chapter 1]](http://repository.uph.edu/style/images/fileicons/text.png)
Chapter1_BRAIN HARLAN.pdf
Available under License Creative Commons Attribution Non-commercial Share Alike.
Download (262kB)
![Chapter 2 [thumbnail of Chapter 2]](http://repository.uph.edu/style/images/fileicons/text.png)
Chapter2_BRAIN HARLAN.pdf
Restricted to Registered users only
Available under License Creative Commons Attribution Non-commercial Share Alike.
Download (555kB)
![Chapter 3 [thumbnail of Chapter 3]](http://repository.uph.edu/style/images/fileicons/text.png)
Chapter3_BRAIN HARLAN.pdf
Restricted to Registered users only
Available under License Creative Commons Attribution Non-commercial Share Alike.
Download (449kB)
![Chapter 4 [thumbnail of Chapter 4]](http://repository.uph.edu/style/images/fileicons/text.png)
Chapter4_BRAIN HARLAN.pdf
Restricted to Registered users only
Available under License Creative Commons Attribution Non-commercial Share Alike.
Download (411kB)
![Chapter 5 [thumbnail of Chapter 5]](http://repository.uph.edu/style/images/fileicons/text.png)
Chapter5_BRAIN HARLAN.pdf
Restricted to Registered users only
Available under License Creative Commons Attribution Non-commercial Share Alike.
Download (116kB)
![Bibliography [thumbnail of Bibliography]](http://repository.uph.edu/style/images/fileicons/text.png)
Bibliography_BRAIN HARLAN.pdf
Available under License Creative Commons Attribution Non-commercial Share Alike.
Download (196kB)
![Appendices [thumbnail of Appendices]](http://repository.uph.edu/style/images/fileicons/text.png)
Appendices_BRAIN HARLAN.pdf
Restricted to Repository staff only
Available under License Creative Commons Attribution Non-commercial Share Alike.
Download (766kB)
Abstract
Dalam beberapa tahun terakhir, Autonomous Mobile Robot (AMR) mulai lebih sering digunakan di gudang dibandingkan Automated Guided Vehicle (AGV). Untuk mengintegrasikan AMR di gudang, algoritma navigasi perlu diterapkan dengan baik. Algoritma Conflict-Based Search (CBS) yang sebelumnya digunakan pada AGV membutuhkan banyak sumber daya untuk dijalankan di lingkungan tersebut. Oleh karena itu, diperlukan peningkatan algoritma CBS agar lebih cocok untuk digunakan pada AMR. Peningkatan yang disarankan adalah dengan menggunakan Multi Agent Reinforcement Learning (MARL) sebagai bentuk algoritma tingkat rendah di atas pencarian jalur tingkat tinggi CBS, dengan mengurangi batasan CBS dan menerapkannya dalam MARL. Hal ini akan mengubah lingkungan dari berbasis grid menjadi lingkungan kontinu yang lebih sesuai dengan AMR.
Alur penelitian dibagi menjadi 4 fase, di mana fase pertama berfokus pada algoritma tingkat tinggi untuk menemukan jalur di lingkungan berbasis grid menggunakan algoritma CBS dan mengekstrak jalur tersebut sebagai gambar untuk digunakan lebih lanjut dalam lingkungan ruang kontinu. Fase kedua adalah melatih kemampuan agen untuk bernavigasi di lintasan terpisah yang dirancang untuk meniru jalur yang mungkin dihasilkan oleh CBS menggunakan algoritma Proximal Policy Optimization (PPO). Fase ketiga adalah menerapkan lingkungan multi-agen dengan menggunakan lintasan yang dihasilkan dan beberapa agen PPO sebagai Multi Agent PPO (MAPPO). Fase terakhir adalah mengevaluasi dan membandingkan kinerja model dengan algoritma CBS sebagai dasar perbandingan.
Hasil eksperimen menunjukkan bahwa model yang diusulkan memiliki kinerja lebih buruk dibandingkan dengan model dasar CBS pada kasus 5 agen dan 10 agen, tetapi lebih baik untuk kasus 15 agen dan 20 agen. Model yang diusulkan berjalan 3,16 kali lebih lambat pada 5 agen, 1,54 kali lebih lambat pada 10 agen, 1,40 kali lebih cepat pada 15 agen, dan berhasil berjalan pada 20 agen di mana model dasar gagal dijalankan. Kegagalan pada model dasar untuk kasus 20 agent diasumsikan disebabkan oleh ketidakmampuan mesin untuk memproses eksperimen dikarenakan beban komputasi yang besar. / In more recent years, Autonomous Mobile Robot (AMR) has started to be used more often in warehouses as opposed to Automated Guided Vehicle (AGV). To incorporate AMR in warehouses, navigation algorithms is needed to be implemented properly. The previously used Conflict-Based Search (CBS) algorithm from AGV can take a lot of resources to run in the environment, thus there is a need to improve the CBS algorithm to suit for AMR use. The improvements suggested is by using Multi Agent Reinforcement Learning (MARL) as a form of low-level algorithm on top of CBS high-level pathfinding by reducing CBS constraints and applying it in MARL. It would change the environment from a grid-based environment to a continuous environment that are more tailored to AMR.
The research pipeline is divided into 4 phases in which the first phase focuses on high-level algorithm to find paths in grid-based environment by using CBS algorithms and extracts the paths as images to further be used as continuous space environment. The second phase is to train an agent’s capabilities to navigate in a separate track that is built to mimic possible paths CBS can generate by using Proximal Policy Optimization (PPO) algorithm. The third phase is to implement multi agent environment by using the generated track and multiple PPO agents as Multi Agent PPO (MAPPO). The final phase is to evaluate and compare the model performance to the baseline CBS algorithm.
The result of the experiment suggests the proposed model when compared to base CBS model, performs worse on 5 agents and 10 agents cases, and performs better for 15 agents and 20 agents cases. The proposed model runs 3,16 times slower in 5 agents, 1,54 times slower in 10 agents, 1,40 times faster in 15 agents, and runs for 20 agents whereas base CBS model fails to run. Failure for base model 20 agent case is assumed to be caused by the machine’s inability to process the experiment due to high computational load.
Item Type: | Thesis (Masters) |
---|---|
Creators: | Creators NIM Email ORCID Harlan, Brain NIM01679230009 UNSPECIFIED UNSPECIFIED |
Contributors: | Contribution Contributors NIDN/NIDK Email Thesis advisor Tjahyadi, Hendra NIDN0410076901 UNSPECIFIED |
Uncontrolled Keywords: | Navigasi ; Conflict-Based Search ; Multi Agent Reinforcement Learning ; Otomasi |
Subjects: | T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK7885-7895 Computer engineering. Computer hardware |
Divisions: | University Subject > Current > Faculty/School - UPH Karawaci > School of Information Science and Technology > Master of Informatics Current > Faculty/School - UPH Karawaci > School of Information Science and Technology > Master of Informatics |
Depositing User: | Phillips Iman Heri Wahyudi |
Date Deposited: | 28 Feb 2025 04:39 |
Last Modified: | 28 Feb 2025 04:39 |
URI: | http://repository.uph.edu/id/eprint/67434 |