Morphing Based Transfer of Demonstrated Surface Finishing Trajectories to Point Clouds of Similar Objects

Abstract

For production in small lot sizes, typical for SMEs, efficient and intuitive programming of process trajectories or transferring them to similar objects is important, particularly for robotic applications such as polishing where frequent reprogramming is needed. Existing approaches rely on CAD models or large training datasets for enabling trajectory transfer. In contrast, this research explores a novel algorithm that facilitate direct transfer of expert demonstrated trajectories between objects represented as point clouds from 3D scans. To achieve performance even without prior object knowledge, we use a neural network-based non-linear model, typically used in animation, to enable morphing of a source to a target point cloud which inherently transforms the associated process trajectories along with it. We address the primary challenge of differences in point clouds by introduction of a keypoint-driven loss function, which aids the model by preserving the local point distribution and thus ensuring accurate trajectory transformation. This innovation avoids the need for any prior training, offering trajectory generation also for new object classes directly at inference. The performance of this algorithm is benchmarked against traditional interpolation methods, demonstrating its superiority in generating trajectories and applicability in industrial settings. This is a significant step forward in adaptive robotic operations, reducing the dependency on extensive datasets and enhancing the flexibility and efficiency of smart manufacturing systems.