MoDeSuite: Robot Learning Task Suite for Benchmarking Mobile Manipulation with Deformable Objects

Yuying Zhang, Kevin Sebastian Luck, Francesco Verdoja, Ville Kyrki, Joni Pajarinen · 2025

Read MoDeSuite: Robot Learning Task Suite for Benchmarking Mobile Manipulation with Deformable Objects by Yuying Zhang, Kevin Sebastian Luck, Francesco Verdoja, Ville Kyrki, Joni Pajarinen (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2507-21796.txt.

Introduction

Yuying Zhang and co-authors launched MoDeSuite in 2025, the first benchmark suite for robots that move and handle soft objects at once. Mobile manipulation is hard enough with rigid items; cloth and other deformables multiply the difficulty. No standard test set covered the mobile, deformable combination. MoDeSuite supplies eight tasks, each echoing a real application and demanding teamwork between the robot's base and its arm.

Two reinforcement learning and two imitation learning systems were trained as a first reading of the suite. Their struggles in simulation are part of the report. Policies were then carried onto a real Spot robot to probe sim-to-real transfer. The authors expect the suite to open a fresh research lane. Code, tasks, and videos are posted for other labs to build on.

Catalog record

CREATORYuying Zhang, Kevin Sebastian Luck, Francesco Verdoja, Ville Kyrki, Joni Pajarinen
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE37 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2507-21796

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

MoDeSuite: Robot Learning Task Suite for Benchmarking Mobile Manipulation with Deformable Objects Yuying Zhang, Kevin Sebastian Luck, Francesco Verdoja, Ville Kyrki, Joni Pajarinen Abstract Mobile manipulation is a critical capability for robots operating in diverse, real-world environments. However, manipulating deformable objects and materials remains a major challenge for existing robot learning algorithms. While various benchmarks have been proposed to evaluate manipulation strategies with rigid objects, there is still a notable lack of standardized benchmarks that address mobile manipulation tasks involving deformable objects. To address this gap, we introduce MoDeSuite, the first Mobile Manipulation Deformable Object task suite, designed specifically for robot learning. MoDeSuite consists of eight distinct mobile manipulation tasks covering both elastic objects and deformable objects, each presenting a unique challenge inspired by real-world robot applications. Success in these tasks requires effective collaboration between the robot’s base and manipulator, as well as the ability to exploit the deformability of the objects. To evaluate and demonstrate the use of the proposed benchmark, we train two state-of-the-art reinforcement learning algorithms and two imitation learning algorithms, highlighting the difficulties encountered and showing their performance in simulation. Furthermore, we demonstrate the practical relevance of the suite by deploying the trained policies directly into the real world with the Spot robot, showcasing the potential for sim-to-real transfer. We expect that MoDeSuite will open a novel research domain in mobile manipulation involving deformable objects. Find more details, code, and videos at https://sites.google.com/view/modesuite/home . I Introduction Mobile manipulation is a complex robotics challenge, integrating robot navigation and object manipulation. Mastering these abilities enables robots to perform intricate and dynamic tasks, ranging from fetching and placing [ 1 ] and opening doors [ 2 ] to fruit harvestings [ 3 ] and human rescue [ 4 ] in disaster scenarios. Many of these tasks involve manipulating deformable objects, a particularly challenging problem that requires further research into robust and adaptable robotic learning techniques [ 5 ] . Deformable objects introduce further unique challenges for mobile manipulators due to their shape variability, which directly impacts manipulation strategies. These challenges can be categorized based on deformation type: plastic or elastic [ 6 ] . Plastic deformation results in permanent structural changes, requiring robust policies that account for irreversible modifications. In contrast, elastic deformation involves temporary, reversible changes, necessitating precise modeling and real-time control to avoid exceeding the material’s elastic limit. Both types demand advanced perception, planning, and control strategies to ensure reliable task execution. Fig. 1 : MoDeSuite features diverse tasks requiring coordinated navigation and manipulation of elastic and plastic deformable materials using both wheeled and legged robots in constrained environments. Despite advancements in both mobile manipulation and deformable object manipulation, there is currently no standardized benchmark that integrates these two domains. This gap limits the ability to systematically develop, evaluate, and compare algorithms across different methodologies. Existing benchmarks primarily focus on either rigid-body mobile [ 7 , 8 ] or static deformable object manipulation [ 9 , 10 ] , leaving mobile deformable manipulation relatively underexplored. Furthermore, real-world experimentation is time-consuming, costly, and difficult to standardize, especially for data-driven approaches like reinforcement learning (RL) and imitation learning (IL). A well-designed simulation-based benchmark could significantly accelerate progress by providing a controlled, reproducible, and scalable testing environment. To address this gap, we introduce MoDeSuite, a standardized task suite specifically designed for mobile deformable manipulation, consisting of eight diverse tasks. As shown in Fig. 1 , MoDeSuite includes two types of mobile manipulators, three types of action spaces, two types of observation spaces, and support for both elastic and plastic object manipulation. Within each task, users can switch between different robots, observation modalities, and action spaces, facilitating flexible experimentation. TABLE I : Comparison of different manipulation frameworks. The check( ✓ ) denotes the presence of the feature. In Supported Dynamics, Plastic denotes plastic deformable objects, including clothes and rope, and Elastic denotes elastic objects such as rubber and foam. In Robotics Platforms, Mobile-M denotes a manipulator with a mobile base, and Legged-M denotes a manipulator with a legged mobile base. Supported Dynamics Robotic Platforms Category Name Physics Engine Plastic Elastic Mobile-M Legged-M Note Mobile Behavior-1k [ 11 ] PhysX5 ✓ ✗ ✓ ✗ Daily Activities AI2THOR [ 12 ] Unity ✗ ✓ ✓ ✗ Indoor Scene TDW Tansport * [ 13 ] PhysX,Flex,Obi ✓ ✓ ✓ ✗ Transport Challenge Habitat [ 14 ] Bullet ✗ ✗ ✓ ✓ Indoor Scene ManiSkill3 [ 15 ] PhysX ✗ ✗ ✓ ✗ Limited Mobile Manipulation Tasks ORBIT [ 16 ] PhysX ✓ ✗ ✓ ✓ FEM,Particle System Deformable DeformableRavens [ 17 ] Bullet ✓ ✓ ✗ ✗ FEM DEDO [ 18 ] Bullet ✓ ✓ ✓ ✗ Particle System DAXBench [ 19 ] DAX ✓ ✓ ✗ ✗ Tabletop, Particle System PlasticineLab [ 20 ] DiffTaiChi ✓ ✓ ✗ ✗ End-effector, Particle System Reform [ 10 ] AGX ✗ ✓ ✗ ✗ End-effector, FEM SoftGym [ 9 ] Flex ✓ ✗ ✗ ✗ Particle System DexGarmentLab [ 21 ] Physx ✓ ✓ ✗ ✗ Particle System, FEM Both MoDeSuite(ours) Physx ✓ ✓ ✓ ✓ Particle System, FEM MoDeSuite is developed within Isaac Lab [ 16 ] and utilizes the high-fidelity simulator Isaac Sim [ 22 ] , enabling efficient training through parallelized environments. Success in these tasks requires agents to exploit object deformability while simultaneously overcoming the dual challenges of navigation and manipulation. To support research in this area, MoDeSuite provides pre-configured models with camera sensors, leveraging the latest advancements in photorealistic rendering and high-fidelity physics simulation. We benchmark four state-of-the-art learning algorithms, two from imitation learning and two from reinforcement learning, and provide a dataset for offline imitation training. To validate the practical significance of our proposed tasks, we implement similar environments in the real world, demonstrating that our benchmark can facilitating sim-to-real transfer. We believe MoDeSuite represents a crucial step forward in the development of mobile deformable manipulation by providing a unified platform for research, benchmarking, and algorithm development. By bridging the gap between mobile manipulation and deformable object interaction, we aim to accelerate progress in both fields. The codebase and detailed installation instructions will be made publicly available upon paper acceptance. II Related Work Mobile Manipulation with Deformables Mobile manipulation involving deformable objects presents significant challenges due to the coordination required between the mobile platform, robotic arms, and the complex dynamics of deformable materials. Most existing approaches that address this problem rely on planning-based methods [ 23 , 24 ] , which often struggle to generalize across diverse tasks and environments. Recently, data-driven methods, particularly imitation learning (IL) and reinforcement learning (RL), have shown promise in mobile manipulation [ 25 , 26 ] and deformable tasks like shape control [ 27 ] and cloth manipulation [ 28 , 29 ] , including mobile deformable scenarios [ 30 ] . However, these methods require either diverse demonstrations or extensive interactions, highlighting the need for simulation environments that offer realistic and diverse tasks to support scalable learning. Mobile Manipulation Benchmark Several existing mobile manipulation benchmarks are designed for specific domains such as underwater [ 31 ] , aerial [ 32 , 33 , 34 ] , assistive [ 35 ] , or rover-based platforms [ 36 ] . More general task suites [

Read the complete text