This thesis presents a multi-modal human demonstration-based skill acquisition framework designed for robotic surface finishing tasks. The conventional iterative robotic teaching methods require repetitive expert demonstrations for each new workpiece,...
This thesis presents a multi-modal human demonstration-based skill acquisition framework designed for robotic surface finishing tasks. The conventional iterative robotic teaching methods require repetitive expert demonstrations for each new workpiece, thus limiting scalability and efficiency. Additionally, imitation learning based solely on positional data often results in suboptimal precision for tasks involving critical contact forces, such as polishing operations.
To address these limitations, we designed a remote teaching tool using a wireless VR tracker integrated with a 6-axis force/torque sensor, enabling synchronous real-time collection of expert's 6-degree-of-freedom pose and contact force data. The collected data underwent mathematical optimization-based calibration to achieve high precision suitable for real robotic operation environments. Additionally, we introduced an artificial neural network (ANN)-based dynamic load compensation (DLC) module to eliminate gravitational and inertial disturbances, yielding more accurate contact force measurements.
Expert demonstrations, recorded iteratively, were temporally aligned via Dynamic Time Warping (DTW), followed by Gaussian Mixture Model and Gaussian Mixture Regression (GMM–GMR) to extract smooth, intention-driven force-motion trajectories. Furthermore, a projection-based generalization strategy was proposed to remap a single planar demonstration onto unknown curved surfaces through centroid alignment, orthogonal projection, and minimal geodesic smoothing, facilitating skill transfer to various geometries without additional demonstrations.
Experimental validations using a UR10 robot for contamination removal tasks demonstrated excellent contamination removal ratios of 100\%, 100\%, and 94.72\% for planar, non-continuous, and edge-shaped surfaces, respectively, outperforming existing methods. Moreover, additional experiments confirmed 100\% contamination removal rates for similar curved surfaces using the proposed generalization strategy, significantly reducing the number of required demonstrations. This research provides a promising solution for robotic automation applications demanding adaptability to diverse geometries and real-time control.