This NAISS project continues and extends the softenable-codesign research line toward deformable and fragile object manipulation, with a particular emphasis on joint morphology–control co-design, imitation learning, and reinforcement learning under soft-body physics. The project is aligned with the European Commission Horizon Europe project SoftEnable, which generalizes rigid-body caging concepts to soft and deformable objects by combining extrinsic geometric constraints with intrinsic material properties. In the present project period, NAISS resources were primarily used to support large-scale parallel simulation and learning in the Genesis simulator, enabling high-fidelity soft-body dynamics for tasks such as gentle grasping, scooping, hanging, and tool-mediated manipulation of fragile objects. A central challenge in deformable-object manipulation is the need for massively parallel environments to stabilize imitation and reinforcement learning under high variance contact dynamics. Soft-body simulation is substantially more computationally demanding than rigid-body simulation, requiring increased GPU usage to maintain sufficient environment throughput. NAISS GPU resources were therefore essential for running hundreds of parallel simulation instances, training policies using stress- and energy-aware rewards, and storing large volumes of rollout data, model checkpoints, and geometry assets. In addition, the project supported new research directions on robotic world models, driven by two newly recruited PhD students in Prof. Florian T. Pokorny’s group and connected to the CloudGripper project (https://cloudgripper.org/ ). World models require GPU-intensive training due to their reliance on large neural architectures (e.g., latent dynamics models, diffusion-based predictors, and vision-centric representations) that learn compact, predictive representations of robot–environment interaction from high-dimensional visual and proprioceptive data. GPUs are critical for training these models efficiently, particularly when scaling across diverse manipulation scenarios and object categories. Overall, NAISS storage and GPU resources were indispensable for enabling reproducible, data-intensive research at the intersection of soft robotics, learning-based manipulation, and model-based reasoning.