Human-Centered Learning for Human-Robot Interaction and Manipulation

Robots struggle with dexterous tasks that humans perform seamlessly in everyday life while working or interacting with others. While Reinforcement Learning can be used to enable complex policies, many tasks cannot be simulated since they might include complex physical interactions between the robot and the environment (including humans). Importantly, this performance discrepancy impedes robotics from operating in the real unstructured world. Learning from humans emerges as an essential paradigm for developing autonomous robotic agents capable of performing complex tasks. Much like humans rarely learn "from zero," instead relying on teachers and demonstrators, robots can benefit greatly from acquiring skills through human input. While the amount of this data is not as abundant in robotic and interactive tasks, as in those relating to vision and language, information can be exploited through multiple modalities such as preferences, demonstrations, or choices. This thesis investigates innovative methods to 1) improve human satisfaction and comfort in Human-Robot Interaction (HRI) tasks while ensuring task performance from the robotic side, and 2) enable robotic agents to solve challenging Manipulation tasks by Learning from Demonstration (LfD) from humans. In the HRI study, handover is examined as an essential interactive task. We observe a realistic handover, potentially susceptible to interruptions or disturbances, proposing a controller with human-like characteristics to address it, and validating it in terms of task performance and human perception of the interaction. Then, an interactive Preference Learning approach is proposed, enabling the robot to tailor the interaction to each user individually by learning from human feedback. Two unique tasks are examined in the study on LfD for Manipulation. In the first one, the insertion of deformable objects is considered by proposing an Imitation Learning approach to solve flexible shoe insole packing, pioneering such an approach. In the second task, the feasibility of LfD approaches in 3D Irregular Object Bin Packing is assessed for the first time. Across these studies, novel Machine Learning-based approaches are proposed to leverage multiple modalities of human input, offering innovative solutions to enhance learning efficiency and effectiveness.