Navigacija mobilnih robota u gužvama pomoću dubokog potpornog učenja

master's thesis

• Title (english): Crowd-aware mobile robot navigation via deep reinforcement learning
• Author: Barbara Pavlović
• Mentor: Marija Seder (mentor)
• Committee member: Ivan Marković (član povjerenstva)
• Committee member: Luka Petrović (član povjerenstva)
• Committee member: Ivan Petrović (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2025-02-25, Croatia

Abstract

U ovom su radu demonstrirana, testirana i uspoređena neka od najpoznatijih rje-
šenja za problem navigacije robota kroz gužve. Svi analizirani modeli temeljeni su na
dubokom potpornom učenju, stoga se njihovo treniranje izvodi principom pokušaja i
pogrešaka. Uspoređuju se izvedbe LM-SARL algoritma u modelima s različitim postup-
cima treniranja i funkcijama nagrade, a radi se i analiza utjecaja intrinzičnih nagrada
na model s istim algoritmom. Analizira se izvedba modela dubokog potpornog... More učenja s
Q-mrežama, ispitivanjem implementacije algoritma SG-D3QN. Svi su modeli trenirani
koristeći jednake eksperimente u okolini kreiranoj programskim jezikom Python, u ko-
jemu su implementirani i ovdje analizirani modeli dubokog učenja. U simulaciji se ko-
risti neholonomski model robota, a testiranja na raznim različitim eksperimentima u
simulaciji izvode se na modelima treniranima u okruženju s ljudskim agentima nevidlji-
vim robotom. Radi se analiza zaglađenosti putanje robota kod svih ispitanih modela, kao
i analiza razine sigurnosti u navigaciji kroz gužvu. Istakuti su uspjesi i kvaliteta ispitanih
modela, no naglašeni su i uočeni nedostaci te mjesto za napredak. Less

Abstract (english)

In this paper, some of the most popular solutions to robot navigation through crowds
have been demonstrated, tested and compared. All analyzed models are based on deep
reinforcement learning and therefore their training is carried out by the principle of trial
and error. Performances of LM-SARL algorithm in models with different training ap-
proaches and reward functions are compared, and the impact of the intrinsic rewards on
models with the same algorithm is also evaluated. The... More performance of deep reinforce-
ment learning model with Q-network is analyzed, by examining the implementation of
SG-D3QN algorithm. All models were trained using the same set of experiments in an
environment created with Python programming language, which was used for the im-
plementation of deep reinforcement learning models as well. A nonholonomic robot
was used in the simulation, and tests with various experiments were performed on the
models trained in an environment with robot invisible to humans. The analysis of path
smoothness and the level of navigation safety is done. The success and quality of the
evaluated models have been highlighted, while the identified shortcomings and room
for improvement have also been emphasized. Less

• Keywords: robot; gužva; navigacija; duboko učenje; duboko potporno učenje; nagrada; izbjegavanje kolizije; prepreke; LM-SARL; SG-D3QN; ICM; RE3
• Keywords (english): robot; crowd; navigation; deep learning; deep reinforcement learning; reward; collision avoidance; obstacles; LM-SARL; SG-D3QN; ICM; RE3
• Language: croatian
• URN:NBN: urn:nbn:hr:168:888574
• Study programme: Title: Computing; Study programme type: university; Study level: graduate; Academic / professional title: magistar/magistra inženjer/inženjerka računarstva (magistar/magistra inženjer/inženjerka računarstva)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Open access
• Created on: 2025-03-12 11:16:09