Probabilistic and Adaptive Methods for Robotic Manipulation and Robot Hand Control under Interaction Uncertainty

This dissertation studies robotic manipulation under interaction uncertainty, including uncertain contact, noisy sensing, human input variability, task reconfiguration, and evaluation variability. Its central claim is that performance improves when sensing, adaptive control, human-in-the-loop programming, probabilistic inference, and evaluation are designed as one coupled pipeline rather than as isolated modules. On the industrial manipulation side, the dissertation presents a stiffness-aware genetic optimization method for elastic interaction, a sensing-integrated dual-arm architecture for wiring harness assembly, and a modular human-in-the-loop design for collaborative robot teaching. The optimization method uses force measurements directly in the fitness loop, avoiding explicit deformable-object modeling in the tested scenarios. The assembly cell demonstrated end-to-end execution over repeated trials, and the connector-programming study showed that users could program the task and later trigger autonomous execution, supporting a proof-of-concept assessment. On the robotic hand side, the dissertation presents probabilistic shared-autonomy methods for grip-force regulation and extends them to hierarchical force-aware skill encoding. By combining sEMG intent estimation, tactile sensing, and Hidden Markov Models, the proposed controllers improved grip-force controllability in the evaluated tasks. The evidence comes from controlled user studies and a pilot condition with one participant with upper-limb loss, while the programming-by-demonstration extension supports autonomous reproduction of demonstrated force-aware skills under limited object variation. The dissertation also introduces a local probabilistic nonlinear latent model for reach-tograsp reconstruction and iHandVR, an immersive platform for prosthetic evaluation. The latent model improved reconstruction quality over the considered global and linear baselines on the studied datasets, and iHandVR combines device-specific simulation, real-time control, and synchronized multimodal acquisition for functional and psychophysical assessment. Taken together, these results support a narrower thesis: explicit treatment of interaction uncertainty can improve manipulation performance and evaluation quality within the studied tasks, datasets, hardware platforms, and participant groups.