LEARNING-BASED ROBOTIC GRASPING AND MANIPULATION SYSTEMS: EVOLUTION OF DATA-DRIVEN APPROACHES AND TASK-INFORMED STRATEGIES

This thesis investigates how to improve data-driven, context-aware robot grasping and manipulation methods when only limited training data are available. It focuses on Learning from Demonstration (LfD), which reduces computational burden by learning from task examples, but may fail to capture all task features due to small datasets. The thesis analyzes LfD approaches by distinguishing between Movement Primitives, which reconstruct demonstrated motions, and Experience Abstraction methods, which extract underlying task behavior. Prominent methods from both categories are experimentally benchmarked using RMSE, R², Information Loss, and time-space complexity to guide the selection of suitable models for specific tasks. The thesis also addresses pre-grasp pose estimation, especially for soft grippers, an area often overlooked in the literature. It proposes a contextualized data-driven model for extracting 6D pre-grasp poses for the Soft Scoop Gripper, using both unsupervised cuboid-based approximation and supervised shape-based feature extraction. GMM and k-means are applied to model and refine grasp-related features. Overall, the thesis shows how LfD-based methods can be tailored to improve grasp planning with limited demonstrations. It highlights future directions such as grasp affordance learning, multimodal imitation, partial-demonstration learning, and hierarchical task modeling.