Exploitation of Similarities in Point Clouds for Simplified Robot Programming by Demonstration

Programming by demonstration empowers robot operators to program robots without coding. This approach finds particular relevance in scenarios involving repetitive assembly processes composed of similar components arranged in patterns. In industrial environments with limited CAD data, like SMEs, point cloud data from scans serves as a practical starting point for further investigations. Recognizing component patterns within point clouds and analyzing demonstrated actions in specific areas are vital for guiding robots in automation tasks. Our work introduces innovative methods designed to identify patterns within point clouds, enabling the efficient analysis of demonstrated actions and their subsequent execution by robotic systems. A primary focus of our approach is the computation of similarities within point clouds, a crucial step in pinpointing regions suitable for repetitive actions. We are particularly intrigued by primitive-based patterns, as these representations align closely with industrial object surfaces. Given the prevalence of geometric representations in industry, recognizing recurring geometries through primitive-based descriptions represents a pertinent research objective. We present the problem within a practical application context and explore the feasibility of comprehensively identifying recurrent geometries using primitive-based features.