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

Francesco Verdoja; Joni Pajarinen; Kevin Sebastian Luck; Ville Kyrki; Yuying Zhang

REVIEW 3 major objections 8 minor 52 references

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

T0 review · 3 major / 8 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A simulation suite gives mobile robots eight deformable-object tasks—and its state-based policies transfer to a real legged robot.

desk verdict Useful benchmark with honest baselines, but the missing success thresholds and unreleased code make the headline numbers unverifiable as written. read the letter →

arxiv 2507.21796 v1 pith:ZGRXH77I submitted 2025-07-29 cs.RO cs.AI

Yuying Zhang , Kevin Sebastian Luck , Francesco Verdoja , Ville Kyrki , Joni Pajarinen This is my paper

classification cs.ROcs.AI

keywords mobilemanipulationdeformableobjectsbenchmarksuiterobotlearningreinforcementimitationsim-to-realtransferelasticandplastic

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

MoDeSuite is a new simulation benchmark for mobile manipulation of deformable objects, built to fill the gap left by benchmarks that handle either rigid-object mobile manipulation or deformable-object manipulation with fixed arms. The suite contains eight tasks—five with elastic rods and belts, three with cloth-like plastic objects—and two robot platforms (wheeled and legged), with both state-based and image-based observations. The authors train two reinforcement learning algorithms and two imitation learning baselines and show that the tasks are solvable yet hard, with the legged platform and vision-based control being the most difficult settings. They also deploy state-based policies trained for a legged robot directly in the real world for two elastic tasks, getting success rates comparable to simulation; an image-based curtain task fails, and the mismatch is traced to a visual domain gap rather than dynamics. The paper's contribution is a standardized, extensible testbed that lets the community compare algorithms on the combined challenge of navigation, manipulation, and deformability.

What carries the argument

The machinery is a layered one: the finite-element method (FEM) simulates elastic bodies with tetrahedral meshes, and position-based dynamics (PBD) simulates cloth-like plastic objects with particles. On top of these physics models, MoDeSuite defines eight task environments with a shared observation tuple (robot state, deformable-object state, and task targets) and a shared base-plus-arm action space, with a discrete action option for data collection. This shared interface is what lets the same RL and IL algorithms run across all tasks and what makes the suite extensible.

What would settle it

Run a MoDeSuite-trained policy on a physical legged robot with an elastic belt whose stiffness or damping is twice the simulated value, and record whether the Place or Drag task still succeeds; if the success rate drops sharply, the sim-to-real claim holds only for the exact simulated material parameters, not for deformable-object transfer in general.

Watch

Extended reading notes

Core claim

MoDeSuite's central claim is that a single simulation suite can cover the combined challenge of mobile manipulation and deformable objects, and that the suite is useful enough to serve as a benchmark. The paper demonstrates the claim in three ways: it defines eight tasks that require coordination between the base and arm and exploit object deformability; it trains four baselines and shows the tasks are solvable but hard, with PPO ahead of SAC and legged platforms harder than wheeled ones; and it transfers state-based policies zero-shot to a physical legged robot on the Place and Drag tasks, with success rates close to simulation. The image-based curtain policy, by contrast, fails on the real robot, and the paper attributes the failure to a visual domain gap rather than to inaccurate simulation of the task dynamics.

Load-bearing premise

The benchmark's usefulness rests on the assumption that the simulator's FEM and PBD models faithfully reproduce real deformable-material behavior; the paper offers only a qualitative trajectory comparison and ten real-world trials per task.

Editorial extensions

If this is right

MoDeSuite provides a common testbed where mobile manipulation algorithms can be compared on deformable-object tasks with elastic and plastic materials, two robot morphologies, and state-based or image-based observations.
State-based policies trained entirely in simulation transfer zero-shot to a real legged robot on the Place and Drag tasks, with success rates of 90–100% across ten real trials.
Image-based imitation policies that succeed in simulation fail in the real world on the Curtain task, and the paper identifies a visual domain gap—not a dynamics mismatch—as the reason.
PPO outperforms SAC on nearly all evaluated tasks, and wheeled platforms are easier than legged ones, indicating where algorithmic and morphological improvements are most needed.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The shared observation interface, which samples a small number of points along an elastic object, could be adapted to rope, multi-link, or multi-object scenarios with minimal change—an extension the paper does not consider.
Because the suite separates robot state, deformable-object state, and task targets, it could serve as a testbed for modular control (navigation first, then manipulation), which the paper does not discuss.
The visual domain gap identified in the Curtain task suggests that adding domain-randomized textures or a small set of real images to the image-based tasks would make the suite a sharper test of sim-to-real generalization for vision-based deformable manipulation.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Short version: this is a useful benchmark paper that fills a real gap—mobile manipulation with deformable objects—but the missing code and unspecified success thresholds make the headline numbers unreproducible as written. The central idea is sound and the experiments are honest; the fix list is clear.

What's actually new: the combination of two mobile platforms (wheeled and legged), eight tasks spanning elastic and plastic deformation, state and image observations, and RL/IL baselines in one suite. The five elastic tasks are new relative to the authors' own DemoBot, and the integration into Isaac Lab/ORBIT with GPU-parallel training is genuinely useful. The baseline results are coherent: PPO generally beats SAC, Spot is harder than Franka, and the BC/Retrieval gap on the plastic tasks is reported without spin. The sim-to-real section is the strongest part—zero-shot deployment on Spot for Place and Drag works with state observations, and the authors are candid that the image-based Curtain policy fails due to visual domain gap. That is honest evidence.

Now the soft spots, in order of severity.

First, the success metrics for Bend, Transport, Drag, and Lift are defined only via distances with no threshold. You cannot compute the success rates in Fig. 3 and Table II from the text. Place is binary, but the other four need a distance-to-Boolean conversion that is never given. Different thresholds could reorder algorithms or change the sim-to-real agreement. For a benchmarking paper, that is load-bearing.

Second, the code is not released—the paper says it will be 'made publicly available upon acceptance.' Reviewers cannot check reward weights, task configurations, or the missing thresholds. This makes the results unverifiable in practice.

Third, the 'first' claim is too strong. Table I lists TDW Transport as supporting deformable objects on a mobile platform; the authors should qualify to 'first suite designed specifically for mobile deformable manipulation.'

Minor: sim-to-real evidence rests on 10 real trials and one qualitative trajectory comparison, so the 'accurate modeling' claim in Section III-A is stronger than the evidence. Success rates over 20 trials lack error bars, so treat them as indicative.

Who this is for: anyone working on mobile manipulation or deformable object manipulation who wants a common testbed. It deserves a serious referee, but only if the authors ship the code and specify all success thresholds. As is, I would not cite it for reproducibility, but I would want to see the revision.

Referee Report

3 major / 8 minor

Summary. MoDeSuite introduces a simulation benchmark for mobile manipulation with deformable objects, built on Isaac Sim/Isaac Lab. It contains eight tasks (five elastic: Place, Bend, Transport, Drag, Lift; three plastic: Uncover, Cover, Curtain) with wheeled (Franka on Ridgeback) and legged (Spot) mobile manipulators, state- and image-based observations, and multiple action spaces. The paper provides PPO and SAC baselines for elastic tasks, BC and retrieval-based IL baselines for plastic tasks, and a small zero-shot sim-to-real deployment on a physical Spot for Place, Drag, and Curtain. The central claims are that MoDeSuite is the first standardized mobile deformable manipulation benchmark, that the tasks are solvable yet challenging, and that state-based policies transfer to the real world with reasonable fidelity.

Significance. If the benchmark is made fully reproducible, MoDeSuite would fill a genuine gap: existing benchmarks cover either mobile manipulation with rigid objects or tabletop deformable manipulation, but rarely both. The suite's breadth—two robot morphologies, elastic and plastic deformables, RL and IL baselines, and a real-robot pilot—is a useful contribution. The paper's strengths include the explicit use of both FEM and PBD, the inclusion of a real Spot deployment, and the attempt to provide baselines across multiple learning paradigms. However, the current manuscript does not yet deliver a reproducible benchmark because key success thresholds and task configurations are not specified, and the sim-to-real evidence rests on small trial counts and qualitative comparisons. The central value of MoDeSuite as a standardized benchmark depends on these gaps being closed.

major comments (3)

[§III-D, Fig. 3, Table II] The success rates reported in Fig. 3 and Table II are not reproducible as written. For Bend, Transport, Drag, and Lift, the text defines success only through distances (e.g., 'distance between the rod's endpoint and the purple target' for Bend, 'distance between the belt's midpoint and the target position' for Lift) without specifying the threshold used to convert a distance into a Boolean success. Since the code is not yet public, a reader cannot recover these thresholds, and different thresholds could reorder PPO versus SAC or change the sim-to-real agreement. The authors must specify the exact success thresholds in the paper or release the evaluation code with the threshold constants.
[§IV-C, Table II, Fig. 5] The sim-to-real claim for state-based policies is supported by only 10 real-world trials per task and no quantitative measure of trajectory fidelity. In Table II, SAC on Drag has 10% simulated success and 0% real success (0 of 10 trials), which is a difference of one trial in simulation; the claim of 'strong sim-to-real alignment' is therefore fragile. Figure 5 provides only a qualitative overlay of simulated and real object trajectories. The manuscript should either report confidence intervals or exact binomial ranges, and should provide a quantitative trajectory error metric (e.g., mean absolute endpoint error over trials) before claiming that the simulation captures the deformable dynamics sufficiently for benchmark use.
[§III-A, §IV-C] The manuscript states that PhysX provides 'accurate modeling of rigid and deformable body physics,' but this assertion is not validated for the deformable materials used. The Curtain task shows a complete sim-to-real failure for both IL policies, attributed to a visual domain gap via a t-SNE plot; however, no quantitative evidence rules out a dynamics mismatch. Because the benchmark's usefulness as a sim-to-real testbed depends on the fidelity of the deformable simulation, the authors should provide a direct validation of the material models (e.g., measured versus simulated deformation under controlled actions) or qualify the claim of high-fidelity physics.

minor comments (8)

[Abstract] The phrase 'the first Mobile Manipulation Deformable Object task suite' is a strong novelty claim; given the related work discussion, the authors should more carefully scope it (e.g., 'to our knowledge') to avoid overclaiming.
[§I] There is a typo: 'facilitating sim-to-real transfer' should be 'facilitating sim-to-real transfer' (the word 'facilitating' is misspelled as 'facilitating').
[§III-A] 'element finite methods' should be 'finite element methods (FEM)'.
[§III-D] The text says 'the task is considered solved if the robot is close to the body target and moves the belt to the other side of the obstacle close to the belt target'; 'close' is undefined, reinforcing the threshold issue raised in the major comments.
[§IV-A] The paper refers to the Franka configuration as 'fixed-base manipulator setting' while the robot is described as a wheeled mobile manipulator; please use consistent terminology.
[§V] 'hyper-mode' is not defined anywhere in the paper; presumably this refers to 'hybrid mode' or a similar concept and should be clarified.
[§III-C] The image encoder is referred to as 'DiNOv2' in the main text and 'DINOv2' in the references; please standardize the capitalization.
[§IV-B] Table III reports success rates over 20 rollouts for IL methods, but the paper does not state whether the same seed/evaluation protocol is used as for the RL results; please specify the evaluation protocol exactly.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the benchmark claims rest on external evaluation and zero-shot sim-to-real transfer, not on fitted predictions or self-citation chains.

full rationale

MoDeSuite is a benchmark and evaluation paper, not a derivation of predicted quantities from fitted inputs. The central claims are that the eight tasks are learnable and that policies trained in simulation transfer zero-shot to a real Spot robot. These claims are supported by independent evaluations: reinforcement learning success rates in Fig. 3 and Tables II-III, imitation learning rollouts, and real-world deployments with no fine-tuning, using state observations from OptiTrack and images from a RealSense camera. No parameter is fitted to a subset of data and then reported as a prediction of a closely related quantity. The only self-citation is DemoBot [30], which is credited as the origin of the three plastic tasks and the retrieval baseline; those tasks are re-described and re-evaluated independently in this paper, and the retrieval method is itself a baseline being compared rather than the source of the benchmark's validity. The FEM/PBD fidelity claim is an assertion about the simulator, checked qualitatively through trajectory comparison and sim-to-real behavior, not a conclusion derived from MoDeSuite's own definitions. The missing distance-to-Boolean thresholds for Bend, Transport, Drag, and Lift success metrics are a reproducibility concern, not circular reasoning: they do not make any reported result equivalent to an input by construction. No uniqueness theorem is imported, no ansatz is smuggled in via citation, and no known result is merely renamed. Therefore the paper shows no significant circularity.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central contribution is an environment, so most premises are domain assumptions about simulation and observation fidelity. No new physical entities are invented. The free parameters are hand-set reward and evaluation design choices whose values are not reported.

free parameters (3)

Task reward shaping weights = Not specified in paper
Each elastic task reward is a weighted sum of distance terms and a stability penalty (Section III-D); the weights are chosen by hand and are not reported, and they influence learned behavior and reported success.
Success thresholds and distances = Not specified in paper
Success is defined as the endpoint or object being near a target, such as the rod endpoint on the table for Place, but no numeric tolerance is given, so the evaluation criteria are not reproducible as stated.
Task-specific sub-targets = Not specified in paper
The Transport intermediate target and the Drag and Lift body targets are hand-designed auxiliary goals added to shape policy behavior (Section III-D); these affect task difficulty and measured performance.

assumptions (3)

domain assumption Isaac Sim with PhysX FEM and PBD faithfully models deformation and contact dynamics for elastic and plastic objects.
Section III-A asserts 'accurate modeling of rigid and deformable body physics'; the only evidence is qualitative trajectory comparisons in Figure 5, not quantitative validation against real material properties.
domain assumption DINOv2 embeddings are a sufficient image observation for imitation learning in these tasks.
Section III-C and IV-B use the DINOv2 image encoder as the image observation; no comparison to other encoders or raw pixels is provided, and the curtain sim-to-real failure highlights the fragility of this assumption.
domain assumption The pretrained Spot locomotion controller is reliable enough for mobile manipulation evaluation.
Section IV-A uses a pretrained controller for the Spot base, and Section IV-B attributes some imitation learning failures to locomotion controller accuracy, indicating this assumption is load-bearing and only partially verified.

how reviews work

0 comments

Cite this review

Pith. "Pith review of MoDeSuite: Robot Learning Task Suite for Benchmarking Mobile Manipulation with Deformable Objects." pith.science (2026). https://pith.science/paper/ZGRXH77I

@misc{pith2026250721796,
  author       = {Pith},
  title        = {Pith review of: MoDeSuite: Robot Learning Task Suite for Benchmarking Mobile Manipulation with Deformable Objects},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZGRXH77I}},
  note         = {Machine review of arXiv:2507.21796}
}

read the original 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.

Figures

Figures reproduced from arXiv: 2507.21796 by the authors.

**Figure 2.** Overview of MoDeSuite. MoDeSuite, built on the [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** The performance of SAC and PPO algorithms on [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: The Curtain task’s Sim-to-Real visual comparison fea [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

**Figure 5.** Figure 5: Sim-Real comparison for the Drag task shows [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

52 extracted references · 42 canonical work pages

[1]

M 2 diffuser: Diffusion-based trajectory optimization for mobile manipulation in 3d scenes,

S. Yan et al. , “M 2 diffuser: Diffusion-based trajectory optimization for mobile manipulation in 3d scenes,” IEEE Transactions on Pattern Analysis and Machine Intelligence , 2025

work page 2025
[2]

Learning to open and traverse doors with a legged manipulator,

M. Zhang, Y . Ma, T. Miki, and M. Hutter, “Learning to open and traverse doors with a legged manipulator,” arXiv preprint arXiv:2409.04882, 2024

arXiv 2024
[3]

Cherry-picking with reinforcement learning

Y . Zhang, L. Ke, A. Deshpande, and A. Gupta, “Cherry-picking with reinforcement learning.”

work page
[4]

Real-time dynamic ges- ture recognition for human-robot collaboration in rescue operations,

S. Dewangan, V . K. Origanti, and F. Kirchner, “Real-time dynamic ges- ture recognition for human-robot collaboration in rescue operations,” in 2024 IEEE International Symposium on Safety Security Rescue Robotics (SSRR). IEEE, 2024, pp. 229–236

work page 2024
[5]

Motion planning for mobile manipulators—a systematic review,

T. Sandakalum and M. H. Ang Jr, “Motion planning for mobile manipulators—a systematic review,” Machines, vol. 10, no. 2, p. 97, 2022

work page 2022
[6]

Callister and D

W. Callister and D. Rethwisch, Materials Science and Engineering: An Introduction , ser. Wiley Plus Products. Wiley, 2008. [Online]. Available: https://books.google.fi/books?id=xNA3OwAACAAJ

work page 2008
[7]

ManipulaTHOR: A framework for visual object manipulation,

K. Ehsani et al. , “ManipulaTHOR: A framework for visual object manipulation,” in 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) . IEEE, pp. 4495–4504. [Online]. Available: https://ieeexplore.ieee.org/document/9578091/

work page arXiv 2021
[8]

robosuite: A modular simulation framework and benchmark for robot learning

Y . Zhu et al. , “robosuite: A modular simulation framework and benchmark for robot learning.” [Online]. Available: http: //arxiv.org/abs/2009.12293

arXiv 2009

Show all 52 references

[9]

SoftGym: Benchmarking deep reinforcement learning for deformable object manipulation

X. Lin, Y . Wang, J. Olkin, and D. Held, “SoftGym: Benchmarking deep reinforcement learning for deformable object manipulation.”
[10]

Reform: A robot learning sandbox for deformable linear object manipulation,

R. Laezza, R. Gieselmann, F. T. Pokorny, and Y . Karayiannidis, “Reform: A robot learning sandbox for deformable linear object manipulation,” in 2021 IEEE International Conference on Robotics and Automation (ICRA) . IEEE, 2021, pp. 4717–4723

2021
[11]

Behavior-1k: A benchmark for embodied ai with 1,000 everyday activities and realistic simulation,

C. Li et al., “Behavior-1k: A benchmark for embodied ai with 1,000 everyday activities and realistic simulation,” in Conference on Robot Learning. PMLR, 2023, pp. 80–93

2023
[12]

Ai2-thor: An interactive 3d environment for visual ai,

E. Kolve et al. , “Ai2-thor: An interactive 3d environment for visual ai,” arXiv preprint arXiv:1712.05474 , 2017

2017 arXiv
[13]

The threedworld transport challenge: A visually guided task-and-motion planning benchmark towards physically real- istic embodied ai,

C. Gan et al. , “The threedworld transport challenge: A visually guided task-and-motion planning benchmark towards physically real- istic embodied ai,” in 2022 International conference on robotics and automation (ICRA). IEEE, 2022, pp. 8847–8854

2022
[14]

Habitat 3.0: A co-habitat for humans, avatars and robots,

X. Puig et al. , “Habitat 3.0: A co-habitat for humans, avatars and robots,” 2023

2023
[15]

Maniskill3: Gpu parallelized robotics simulation and rendering for generalizable embodied ai,

S. Tao et al. , “Maniskill3: Gpu parallelized robotics simulation and rendering for generalizable embodied ai,” Robotics: Science and Systems, 2025

2025
[16]

ORBIT: A unified simulation framework for interactive robot learning environments

M. Mittal et al. , “ORBIT: A unified simulation framework for interactive robot learning environments.” [Online]. Available: http://arxiv.org/abs/2301.04195

arXiv
[17]

Learning to Rearrange Deformable Cables, Fabrics, and Bags with Goal-Conditioned Transporter Networks,

D. Seita et al. , “Learning to Rearrange Deformable Cables, Fabrics, and Bags with Goal-Conditioned Transporter Networks,” in IEEE International Conference on Robotics and Automation (ICRA) , 2021

2021
[18]

Dynamic environments with deformable objects,

R. Antonova, P. Shi, H. Yin, Z. Weng, and D. K. Jensfelt, “Dynamic environments with deformable objects,” in Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2), 2021

2021
[19]

Daxbench: Benchmarking deformable object ma- nipulation with differentiable physics,

S. Chen et al. , “Daxbench: Benchmarking deformable object ma- nipulation with differentiable physics,” in The Eleventh International Conference on Learning Representations
[20]

Plasticinelab: A soft-body manipulation benchmark with differentiable physics,

Z. Huang et al., “Plasticinelab: A soft-body manipulation benchmark with differentiable physics,” in International Conference on Learning Representations, 2021. [Online]. Available: https://openreview.net/ forum?id=xCcdBRQEDW

2021
[21]

Dexgarmentlab: Dexterous garment manip- ulation environment with generalizable policy,

Y . Wang et al. , “Dexgarmentlab: Dexterous garment manip- ulation environment with generalizable policy,” arXiv preprint arXiv:2505.11032, 2025

2025
[22]

Nvidia isaac sim,

Nvidia, “Nvidia isaac sim,” https://developer.nvidia.com/isaac-sim, May 2022

2022
[23]

Collaborative object manipulation through indirect control of a deformable sheet by a mobile robotic team,

K. Hunte and J. Yi, “Collaborative object manipulation through indirect control of a deformable sheet by a mobile robotic team,” in 2019 IEEE 15th International Conference on Automation Science and Engineering (CASE). IEEE, 2019, pp. 1463–1468

2019
[24]

Planning and control for deformable linear object manipulation,

B. Aksoy and J. Wen, “Planning and control for deformable linear object manipulation,” arXiv preprint arXiv:2503.04007 , 2025

2025
[25]

Combining learning-based locomotion policy with model-based manipulation for legged mobile manipulators,

Y . Ma, F. Farshidian, T. Miki, J. Lee, and M. Hutter, “Combining learning-based locomotion policy with model-based manipulation for legged mobile manipulators,” vol. 7, no. 2, pp. 2377–2384
[26]

Learning kinematic feasibility for mobile manipulation through deep reinforcement learn- ing,

D. Honerkamp, T. Welschehold, and A. Valada, “Learning kinematic feasibility for mobile manipulation through deep reinforcement learn- ing,” vol. 6, no. 4, pp. 6289–6296
[27]

Learning shape control of elastoplas- tic deformable linear objects,

R. Laezza and Y . Karayiannidis, “Learning shape control of elastoplas- tic deformable linear objects,” in 2021 IEEE International Conference on Robotics and Automation (ICRA) , pp. 4438–4444

2021
[28]

Cloth manipulation using random-forest-based imitation learning,

B. Jia, Z. Pan, Z. Hu, J. Pan, and D. Manocha, “Cloth manipulation using random-forest-based imitation learning,” IEEE Robotics and Automation Letters, vol. 4, no. 2, pp. 2086–2093, 2019

2019
[29]

Learning deformable object manipulation from expert demonstra- tions,

G. Salhotra, I.-C. A. Liu, M. Dominguez-Kuhne, and G. S. Sukhatme, “Learning deformable object manipulation from expert demonstra- tions,” IEEE Robotics and Automation Letters, vol. 7, no. 4, pp. 8775– 8782, 2022

2022
[30]

Demobot: Deformable mobile manipulation with vision-based sub-goal retrieval,

Y . Zhang, W. Yang, G. Sivasubramanian, and J. Pajarinen, “Demobot: Deformable mobile manipulation with vision-based sub-goal retrieval,”
[31]

Exploring 3-d reconstruction techniques: A bench- marking tool for underwater robotics,

J. Perez et al. , “Exploring 3-d reconstruction techniques: A bench- marking tool for underwater robotics,” vol. 22, no. 3, pp. 85–95, conference Name: IEEE Robotics & Automation Magazine
[32]

Benchmarks for aerial manipulation,

A. Suarez, V . M. Vega, M. Fernandez, G. Heredia, and A. Ollero, “Benchmarks for aerial manipulation,” vol. 5, no. 2, pp. 2650–2657, conference Name: IEEE Robotics and Automation Letters
[33]

Aerial gym simulator: A framework for highly parallelized simulation of aerial robots,

M. Kulkarni, W. Rehberg, and K. Alexis, “Aerial gym simulator: A framework for highly parallelized simulation of aerial robots,” IEEE Robotics and Automation Letters , 2025

2025
[34]

Omnidrones: An efficient and flexible platform for reinforcement learning in drone control,

B. Xu, F. Gao, C. Yu, R. Zhang, Y . Wu, and Y . Wang, “Omnidrones: An efficient and flexible platform for reinforcement learning in drone control,” IEEE Robotics and Automation Letters , vol. 9, no. 3, pp. 2838–2844, 2024

2024
[35]

Assistive gym: A physics simulation framework for assistive robotics,

Z. Erickson, V . Gangaram, A. Kapusta, C. K. Liu, and C. C. Kemp, “Assistive gym: A physics simulation framework for assistive robotics,” in 2020 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2020, pp. 10 169–10 176

2020
[36]

Rlroverlab: An advanced reinforce- ment learning suite for planetary rover simulation and training,

A. B. Mortensen and S. Bøgh, “Rlroverlab: An advanced reinforce- ment learning suite for planetary rover simulation and training,” in 2024 International Conference on Space Robotics (iSpaRo) . IEEE, 2024, pp. 273–277

2024
[37]

Behavior robot suite: Streamlining real-world whole- body manipulation for everyday household activities,

Y . Jiang et al., “Behavior robot suite: Streamlining real-world whole- body manipulation for everyday household activities,” arXiv preprint arXiv:2503.05652, 2025

2025 arXiv
[38]

Empowering embodied manipulation: A bimanual- mobile robot manipulation dataset for household tasks,

T. Zhang et al. , “Empowering embodied manipulation: A bimanual- mobile robot manipulation dataset for household tasks,” arXiv preprint arXiv:2405.18860, 2024

2024 arXiv
[39]

Habitat 2.0: Training home assistants to rearrange their habitat,

A. Szot et al. , “Habitat 2.0: Training home assistants to rearrange their habitat,” in Advances in Neural Information Processing Systems (NeurIPS), 2021

2021
[40]

Benchmarking and simulating bimanual robot shoe lacing,

H. Luo and Y . Demiris, “Benchmarking and simulating bimanual robot shoe lacing,” IEEE Robotics and Automation Letters , vol. 9, no. 10, pp. 8202–8209, 2024

2024
[41]

Orbit-surgical: An open-simulation framework for learning surgical augmented dexterity,

Q. Yu et al. , “Orbit-surgical: An open-simulation framework for learning surgical augmented dexterity,” in 2024 IEEE International Conference on Robotics and Automation (ICRA) . IEEE, 2024, pp. 15 509–15 516

2024
[42]

Modeling, learning, perception, and control methods for deformable object manipulation,

H. Yin, A. Varava, and D. Kragic, “Modeling, learning, perception, and control methods for deformable object manipulation,” Science Robotics, vol. 6, no. 54, p. eabd8803, 2021

2021
[43]

Available: \url{https://www.franka.de/}

“Franka.” [Online]. Available: \url{https://www.franka.de/}
[44]

Ridgeback,

“Ridgeback,” https://clearpathrobotics.com/ ridgeback-indoor-robot-platform/
[45]

“Spot,” https://www.bostondynamics.com/products/spot
[46]

Spot arm,

“Spot arm,” https://www.bostondynamics.com/products/spot/arm
[47]

Dinov2: Learning robust visual features without supervision,

M. Oquab et al. , “Dinov2: Learning robust visual features without supervision,” 2023

2023
[48]

Proximal policy optimization algorithms,

J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, “Proximal policy optimization algorithms,” arXiv preprint arXiv:1707.06347, 2017

2017 arXiv
[49]

Soft actor-critic: Off- policy maximum entropy deep reinforcement learning with a stochastic actor,

T. Haarnoja, A. Zhou, P. Abbeel, and S. Levine, “Soft actor-critic: Off- policy maximum entropy deep reinforcement learning with a stochastic actor,” in International conference on machine learning . PMLR, 2018, pp. 1861–1870

2018
[50]

rl-games: A high-performance framework for reinforcement learning,

D. Makoviichuk and V . Makoviychuk, “rl-games: A high-performance framework for reinforcement learning,” https://github.com/Denys88/rl games, May 2021

2021
[51]

A framework for behavioural cloning

M. Bain and C. Sammut, “A framework for behavioural cloning.” in Machine Intelligence 15 , 1995, pp. 103–129

1995
[2024]

Available: https://arxiv.org/abs/2408.15919

[Online]. Available: https://arxiv.org/abs/2408.15919

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

M 2 diffuser: Diffusion-based trajectory optimization for mobile manipulation in 3d scenes,

S. Yan et al. , “M 2 diffuser: Diffusion-based trajectory optimization for mobile manipulation in 3d scenes,” IEEE Transactions on Pattern Analysis and Machine Intelligence , 2025

work page 2025

[2] [2]

Learning to open and traverse doors with a legged manipulator,

M. Zhang, Y . Ma, T. Miki, and M. Hutter, “Learning to open and traverse doors with a legged manipulator,” arXiv preprint arXiv:2409.04882, 2024

arXiv 2024

[3] [3]

Cherry-picking with reinforcement learning

Y . Zhang, L. Ke, A. Deshpande, and A. Gupta, “Cherry-picking with reinforcement learning.”

work page

[4] [4]

Real-time dynamic ges- ture recognition for human-robot collaboration in rescue operations,

S. Dewangan, V . K. Origanti, and F. Kirchner, “Real-time dynamic ges- ture recognition for human-robot collaboration in rescue operations,” in 2024 IEEE International Symposium on Safety Security Rescue Robotics (SSRR). IEEE, 2024, pp. 229–236

work page 2024

[5] [5]

Motion planning for mobile manipulators—a systematic review,

T. Sandakalum and M. H. Ang Jr, “Motion planning for mobile manipulators—a systematic review,” Machines, vol. 10, no. 2, p. 97, 2022

work page 2022

[6] [6]

Callister and D

W. Callister and D. Rethwisch, Materials Science and Engineering: An Introduction , ser. Wiley Plus Products. Wiley, 2008. [Online]. Available: https://books.google.fi/books?id=xNA3OwAACAAJ

work page 2008

[7] [7]

ManipulaTHOR: A framework for visual object manipulation,

K. Ehsani et al. , “ManipulaTHOR: A framework for visual object manipulation,” in 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) . IEEE, pp. 4495–4504. [Online]. Available: https://ieeexplore.ieee.org/document/9578091/

work page arXiv 2021

[8] [8]

robosuite: A modular simulation framework and benchmark for robot learning

Y . Zhu et al. , “robosuite: A modular simulation framework and benchmark for robot learning.” [Online]. Available: http: //arxiv.org/abs/2009.12293

arXiv 2009