The European Commission project SoftEnable extends the concept of rigid body caging to soft, deformable objects by integrating both extrinsic and intrinsic constraints. This enables the development of robust manipulation primitives that can handle perturbations effectively. This request is for a Storage accompanying an ongoing Alvis compute project proposal (NAISS 2025/5-43 (C3SE)), where I am a proxy and my supervisor, Florian Pokorny, is the Principal Investigator. The default storage allocation (500G) is insufficient for our needs. Our previous project (NAISS 2024/5-401 (C3SE)), where I am also a proxy, is ending shortly and has also proven insufficient in terms of storage capacity. While a follow-up compute proposal is being prepared, continued storage is critical for ensuring uninterrupted access to datasets and trained models. This project involves AI/ML-driven optimization for robotic manipulation and consists of two key components: 1. Bilevel Optimization for Tool/End-Effector Design (part of SoftEnable): A Bayesian optimization framework proposes candidate designs in the outer loop. Reinforcement Learning (RL) agents, trained using Alvis compute resources, evaluate manipulation policies for these designs. The RL agent processes RGB images of rigid and deformable objects and outputs corresponding tool/gripper actions. The approach focuses on robust manipulation strategies (e.g., pushing, scooping) with manipulation robustness metrics as reward signals. 2. Diffusion Models for Imitation Learning & Skill Chaining (new project): Diffusion models (DDPMs) compute escape energy for robust long-horizon manipulation planning. DDPMs also learn primitive manipulation skills (e.g., pushing, picking, placing) from expert demonstrations. These learned skills are chained using Task and Motion Planning (TAMP) algorithms such as PDDLStream to enable complex manipulation tasks. A curated dataset of escape paths is used to train DDPMs, enabling fast, multi-modal inference of manipulation plans that outperform traditional sampling-based methods. Both projects involve computationally intensive AI/ML training requiring high-performance GPUs. The requested storage is necessary to maintain continuity in our work while we finalize the next compute proposal.