Skill learning and task composition from human demonstrations for a collaborative manipulator

Human-robot collaboration (HRC) in the modern industry requires the employment of manipulators that can acquire and reuse skills in a easy way and without domain-specific knowledge. Learning from Demonstration (LfD) offers a practical way to do so, however, real deployments still face some caveats; such as turning raw demonstrations into reliable low-level controllers on hardware, re-parameterizing skills to new object/goal poses. This thesis presents a unified LfD pipeline implementation for the UFACTORY xArm6, a 6DOF collaborative robot (cobo) that allows learning of reusable motor skills from human kinesthetic demonstrations. Such demonstrations can be used to plan more complex manipulation task. The pipeline follows the canonical stages -- demonstration data acquisition, motion encoding, execution, and refinement -- and is developed with three learning methods for low-level skills: Behavioral Cloning (BC), Dynamic Movement Primitives (DMP), and Gaussian Mixture Models with Gaussian Mixture Regression (GMM-GMR). The Task-dependent parameters are retrieved from a vision subsystem based on RGB-D RealSense D435 camera, enabling skill adaptation to unseen situations without retraining. On the real robot the execution uses the vendor SDK for fine-grained control which in turn allows to tackle the Sim2Real gap; in simulation, trajectories are executed through MoveIt environment for rapid checking. The pipeline is validated on four low-level skills: PICK, PLACE, POUR, and SHAKE. Moreover, it is also tested on high-level tasks that organize them into plans, including pick-and-place, object collection, single-drink pouring, and multi-ingredient mixing. Furthermore, we assume that skills (eventually conditioned) sequencing is user-defined. The overall performance is assessed with geometry and objective fulfillment metrics (Cartesian/orientation RMSE, Hausdorff distance, endpoint error), motion quality (jerk), and task-level success rates, highlighting trade-offs between learning approaches (e.g., GMM-GMR's accuracy on discrete motions and DMPs' convenient goal re-targeting). The contributions are: (i) a modular, end-to-end LfD pipeline for the cobot that goes from kinesthetic data capture to real-robot execution; (ii) a comparative implementation of BC, DMP, and GMM-GMR for skill learning; (iii) a vision-guided parameterization layer for skill reuse across poses; and (iv) task-level controllers that include learned primitives into reliable plans. Collectively, these results demonstrate that few-demo LfD can deliver accurate, adaptable, and maintainable behaviors for collaborative manipulation, reducing integration effort while preserving generality.