A Reinforcement Learning Approach to View Planning for Automated Inspection Tasks.

Area of Science:

• Robotics
• Artificial Intelligence
• Computer-Aided Design (CAD)

Background:

• Manual inspection in flexible manufacturing is costly and unreliable.
• Reinforcement Learning (RL) shows potential for intelligent automation in manufacturing and inspection.
• Automated inspection systems are needed for flexible production with small lot sizes.

Purpose of the Study:

• To develop an RL-based framework for determining optimal sensor view poses for arbitrary 3D CAD models.
• To create an expandable and comparable benchmark for RL algorithms in robotic inspection tasks.
• To integrate the framework with Robot Operating System (ROS) for versatile robot and sensor deployment.

Main Methods:

• Utilizing Reinforcement Learning (RL) to learn optimal sensor view poses from 3D CAD data.
• Extending open-source libraries and integrating with Robot Operating System (ROS).
• Benchmarking using OpenAI Gym and Baselines for algorithm comparison and validation.

Main Results:

• Achieved a coverage ratio of up to 0.8 in experimental scenarios.
• Demonstrated the framework's functionality and potential impact through RL algorithm comparisons.
• Successfully integrated with ROS for flexible robot and sensor deployment.

Conclusions:

• The proposed RL framework effectively determines high-quality sensor view poses for automated workpiece inspection.
• The system offers a promising, expandable, and comparable benchmark for RL in robotic manufacturing.
• Public release of the project will foster further research and development in automated inspection.