Please use this identifier to cite or link to this item:
Title: Reinforcement learning-based method 3D motion planning for a 3D inspection task using prior viewpoint information
Authors: Chee, Yeng Sung
Keywords: Engineering::Mechanical engineering
Issue Date: 2022
Publisher: Nanyang Technological University
Source: Chee, Y. S. (2022). Reinforcement learning-based method 3D motion planning for a 3D inspection task using prior viewpoint information. Final Year Project (FYP), Nanyang Technological University, Singapore.
Abstract: Surface inspection and shape reconstruction is a common application in the factory production line. In a robotic inspection task, generating an optimal collision-free trajectory while meeting the coverage requirement of the target object is challenging because there are inspection cost and traveling cost that needs to be optimized. This paper is based on previous work that proposed a computational framework for automatic online path generation for robotic inspection via coverage planning and reinforcement learning-based approach. The online processing stage is utilising Monte Carlo Tree Search (MCTS) with the formulation of Markov Decision Process (MDP). However, a proposed visibility modelling and approximation that considers the presence of an obstacle is introduced to tackle the issue of obstructed view of the target object seen from a 3D camera. The proposed method compares the distance traveled by the same camera rays with a minimum threshold in two scenarios: target object with and without an obstacle to obtain a more realistic visibility of the viewpoint. The proposed MCTS in this final year project modifies the tree policy and default policy. The selection process by the tree policy is based on the reward of MCTS instead of the costs. The definitions of inspection cost and traveling cost are also modified to the number of covered surfaces (the uncovered surfaces of the selected viewpoints) of unselected viewpoint, and the trajectory length respectively. Instead of choosing the viewpoints greedily in the default policy, the average of sum of probabilities of the 2 costs is used to choose an action. After the parameters of MCTS are tuned, the proposed MCTS algorithm will be experimented on two scenes with and without obstacle and on various target objects.
Schools: School of Mechanical and Aerospace Engineering 
Organisations: A*STAR Institute of Infocomm Research (I2R)
Fulltext Permission: restricted
Fulltext Availability: With Fulltext
Appears in Collections:MAE Student Reports (FYP/IA/PA/PI)

Files in This Item:
File Description SizeFormat 
  Restricted Access
2.55 MBAdobe PDFView/Open

Page view(s)

Updated on Dec 11, 2023


Updated on Dec 11, 2023

Google ScholarTM


Items in DR-NTU are protected by copyright, with all rights reserved, unless otherwise indicated.