Design, control and computational strategies for unstructured robotic applications

Modern robotics has moved from tightly scripted, structured workcells to unstructured, open-world settings. This shift replaces fixed fixtures and repeatable parts with deformable objects, occlusions, and changing contacts. As a result, manipulation now hinges on perception-driven control, uncertainty handling, and learning from data. We embrace this transition and show how combining model-based priors with adaptive policies enables reliable performance beyond the lab. The main contribution focuses on deformable linear objects manipulation: industrial appli-cations are developed for wire harness assembly, hose manipulation, and cable storage using a multimodal system that combines vision and tactile sensing. Then, a tactile approach is pro-posed, where tactile readings are given physical meaning through a Hertzian contact model with Bayesian parameter estimation, enabling reliable force regulation and alignment during grasping. This work explores 3D-printing applications integrated with Wire Arc Additive Manufactur-ing . A 14-degree-of-freedom climbing robot is developed: it can anchor to the printed structure with variable geometries to preserve process continuity. A library for capability maps is developed, encompassing maps of reachability, manipula-bility, and collision that directly feed motion planning. In unstructured environments, bridging the Sim2Real gap is essential to transition from laboratory tuning to stable production. We adopt parameter identification, targeted random-ization, and progressive validation to make simulation predictive within known limits. The final work to guide robot in unstructured environments converges into a Task-Priority control framework that actively manages robot limits and constraints while leveraging Robotic Distance Functions, a differentiable distance representations of geometry, to achieve gradient-consistent collision avoidance. In unstructured scenarios, this framework becomes crucial: dy-namic task activation and coherent gradient use make planning mistakes less likely and execu-tion more robust.