Enhancing Visual Inspection in Remanufacturing: A Reinforcement Learning Approach with Integrated Robot Simulation
Research output: Journal contributions › Conference article in journal › Research › peer-review
Standard
In: Procedia CIRP, Vol. 134, 2025, p. 939-944.
Research output: Journal contributions › Conference article in journal › Research › peer-review
Harvard
APA
Vancouver
Bibtex
}
RIS
TY - JOUR
T1 - Enhancing Visual Inspection in Remanufacturing
T2 - 58th CIRP Conference on Manufacturing Systems, CMS 2025
AU - Koch, Dominik
AU - Kaiser, Jan Philipp
AU - Stamer, Florian
AU - Stark, Rainer
AU - Lanza, Gisela
N1 - Publisher Copyright: © 2025 Elsevier B.V.. All rights reserved.
PY - 2025
Y1 - 2025
N2 - This paper presents an approach for solving the View Planning Problem (VPP) in robotic inspection using Reinforcement Learning (RL). Building on a prior framework, this work takes a significant step forward by integrating a detailed robotic simulation environment with essential modules for trajectory and reachability planning. This allows for the development of an RL agent that not only selects adaptive viewpoints but also considers kinematic constraints and collision-free paths, which are crucial for practical, real-world inspections. The study specifically targets the initial inspection of returned products with high variability, demonstrating the feasibility of RL to manage complex tasks in remanufacturing. The RL-based solution is evaluated using Soft Actor-Critic (SAC) and Proximal Policy Optimization algorithms, with SAC showing superior performance. The learned strategies were validated on a real inspection station, showing the capability of using RL based inspection strategies. This research offers a robust, adaptable solution for inspection challenges, bridging the gap between theoretical models and application-ready inspection systems.
AB - This paper presents an approach for solving the View Planning Problem (VPP) in robotic inspection using Reinforcement Learning (RL). Building on a prior framework, this work takes a significant step forward by integrating a detailed robotic simulation environment with essential modules for trajectory and reachability planning. This allows for the development of an RL agent that not only selects adaptive viewpoints but also considers kinematic constraints and collision-free paths, which are crucial for practical, real-world inspections. The study specifically targets the initial inspection of returned products with high variability, demonstrating the feasibility of RL to manage complex tasks in remanufacturing. The RL-based solution is evaluated using Soft Actor-Critic (SAC) and Proximal Policy Optimization algorithms, with SAC showing superior performance. The learned strategies were validated on a real inspection station, showing the capability of using RL based inspection strategies. This research offers a robust, adaptable solution for inspection challenges, bridging the gap between theoretical models and application-ready inspection systems.
KW - Automation
KW - Reinforcement Learning
KW - Robotics
KW - Visual inspection
KW - Engineering
UR - http://www.scopus.com/inward/record.url?scp=105009412682&partnerID=8YFLogxK
U2 - 10.1016/j.procir.2025.02.228
DO - 10.1016/j.procir.2025.02.228
M3 - Conference article in journal
AN - SCOPUS:105009412682
VL - 134
SP - 939
EP - 944
JO - Procedia CIRP
JF - Procedia CIRP
SN - 2212-8271
Y2 - 13 April 2025 through 16 April 2025
ER -