Data-Efficient Representation Learning for Grasping and Manipulation

Doctoral thesis, 2026

General-purpose robotic manipulation requires adaptability to diverse environments. To achieve this, robots need effective representations that allow them to understand their environment and perform correct actions. Machine learning provides a strong basis for designing and teaching robots such representations. While machine learning has driven remarkable progress in areas such as computer vision and natural language processing, its application to robotics is constrained by the high cost and complexity of collecting real-world training data. In robotics, collecting real-world data requires physical interaction with the environment and often involves expert-designed task setups, human demonstrations, careful supervision, and extensive data curation, making the process expensive and difficult to scale. This creates a critical need for data-efficient learning approaches. This thesis presents a framework for data-efficient robotic manipulation through structured, task-aware representations. We investigate part-based representations for transferring grasp information, scene representations that capture essential structure for motion modeling, and reactive motion models that enable adaptive execution. Collectively, these contributions demonstrate that embedding task-relevant structure into robot representations provides a foundation for teaching robots diverse manipulation skills, enabling new capabilities, and substantially improving data efficiency.