Caro, Steven2026-06-232026-06-232026-06-232026-05-29https://hdl.handle.net/10012/23657Nonprehensile manipulation, i.e. interaction without grasping, is a fundamental capability for mobile robots operating in unstructured environments, yet it remains a challenging control problem due to complex contact dynamics and under-actuated physics. Progress in this domain has been further hindered by the lack of standardized evaluation frameworks, leading to fragmented research efforts. This thesis addresses these gaps through two primary contributions: a unified benchmarking suite and a novel hierarchical control architecture. First, we introduce Bench-Push, a comprehensive benchmark designed specifically for pushing-based mobile robot tasks. Unlike existing benchmarks that penalize environment interaction, Bench-Push provides diverse environments — ranging from navigation-centric mazes to manipulation-centric delivery tasks — and introduces novel metrics to quantify the trade-off between task efficiency and interaction effort. We validate the framework by evaluating state-of-the-art baselines and demonstrating successful zero-shot transfer to a physical robot. Second, we propose the Hierarchical Reinforcement Learning - Diffusion Policy (HeRD), a hybrid architecture designed to solve long-horizon manipulation tasks. We identify that while Reinforcement Learning (RL) excels at strategic decision-making, it struggles to learn precise low-level contact dynamics. Conversely, Generative Diffusion Models synthesize smooth, context-aware trajectories, but lack high-level planning capabilities. HeRD bridges this gap by decoupling the control hierarchy: a high-level RL planner, utilizing a Spatial Action Map action space, selects strategic subgoals, which are then executed by a low-level, goal-conditioned diffusion policy. Extensive experiments in the Box-Delivery task demonstrate that HeRD significantly outperforms both state-of-the-art learning-based methods (SAM) and classical motion planners (Greedy Heuristic, Hierarchical RRT*). HeRD achieves higher success rates and greater interaction efficiency in both simulation and real-world deployments. Furthermore, we demonstrate that HeRD is capable of robust zero-shot generalization to unseen, unstructured clutter, successfully navigating complex environments where classical planners experience catastrophic failure.enmobile manipulationreinforcement learninglearning from demonstrationMaster Thesis