pith. machine review for the scientific record. sign in

arxiv: 2604.25126 · v1 · submitted 2026-04-28 · 💻 cs.RO

Recognition: unknown

HANDFUL: Sequential Grasp-Conditioned Dexterous Manipulation with Resource Awareness

Ethan Foong , Yunshuang Li , Hao Jiang , Gaurav S. Sukhatme , Daniel Seita
Authors on Pith no claims yet

Pith reviewed 2026-05-07 16:14 UTC · model grok-4.3

classification 💻 cs.RO
keywords dexterous manipulationsequential tasksgrasp planningresource awarenessfinger-level rewardscurriculum learningsimulation benchmarkmultifunctional manipulation
0
0 comments X

The pith

Treating fingers as a limited resource during initial grasps improves success on follow-up dexterous manipulation tasks.

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

The paper establishes that sequential dexterous tasks, where a robot must grasp an object and then execute a distinct second action like pushing or pressing while keeping the grasp, require planning initial grasps that conserve fingers rather than maximizing only the first action. A sympathetic reader would care because most existing dexterous manipulation work targets isolated single-skill episodes, yet practical multifunctional use demands maintaining object control across steps without exhausting hand resources. The authors introduce a framework that adds finger-level contact rewards to promote such resource-aware grasps and then trains policies on them through curriculum learning. They support this with a new benchmark containing multiple grasp-conditioned second-subtask objectives and show higher second-subtask success and robustness than a greedy baseline that ignores future finger needs. Real-robot validation on a physical hand further indicates the principle transfers beyond simulation.

Core claim

By modeling finger usage as a limited resource and applying finger-level contact rewards during grasp learning, the framework produces initial grasps that preserve both stability and finger availability; when these grasps are selected via curriculum-based policy learning for downstream subtasks, second-subtask success rates and robustness rise compared with baselines that greedily optimize only the initial grasp.

What carries the argument

Finger-level contact rewards that treat individual fingers as scarce resources during grasp optimization, combined with curriculum-based policy selection to favor grasps compatible with a second subtask.

If this is right

  • Prioritizing resource conservation in the initial grasp directly raises the probability of completing the second subtask without dropping the object.
  • Curriculum learning on resource-aware grasps produces policies that generalize better across pushing, pulling, and pressing objectives under the same grasp-conditioned setup.
  • The same principle yields measurable robustness gains in simulation and transfers to physical execution on a multi-fingered hand.
  • A shared benchmark of sequential tasks provides a concrete testbed for comparing grasp strategies that account for future finger availability.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Longer task sequences beyond two subtasks would likely require extending the reward structure to track remaining finger budgets across multiple steps.
  • The approach could be combined with perception modules so that grasp selection adapts online to object geometry variations not seen in simulation.
  • Explicit resource modeling may reduce the need for frequent re-grasping in real deployments where finger fatigue or contact drift occurs.

Load-bearing premise

Finger-level contact rewards during the first grasp will consistently yield holds that remain stable while leaving the needed fingers free for the second subtask, even across different objects and real-world conditions.

What would settle it

A controlled comparison on the benchmark tasks in which the resource-aware method shows no gain or a loss in second-subtask success rate relative to the greedy initial-grasp baseline, under identical training budgets and object sets.

Figures

Figures reproduced from arXiv: 2604.25126 by Daniel Seita, Ethan Foong, Gaurav S. Sukhatme, Hao Jiang, Yunshuang Li.

Figure 1
Figure 1. Figure 1: HANDFUL enables sequential dexterous manipulation by learning resource-aware grasps that preserve specific fingers for downstream subtasks. For each task, the robot first selects an appropriate initial grasp of the object (a block), and then executes a second subtask using the remaining fingers. We show real-world rollouts of HANDFUL using a LEAP Hand [28] for tasks that involve pushing (top left), pressin… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of HANDFUL. (A, Sec. IV-A): Multiple grasping policies are trained with active and inactive finger constraints to encourage resource-aware grasps that preserve fingers for future manipulation. (B, Sec. IV-B): For each grasp, a second-stage manipulation policy is trained via a multi-stage curriculum, where survivor policies are selected at increasing environment difficulty levels. (C, Sec. IV￾C): S… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of HANDFUL-Bench (Sec. V). The top row shows the starting scenarios for the five tasks. The middle row shows the outcome after grasping the block with resource-aware grasping policies. The bottom row shows successful configurations after performing the second subtask while preserving the initial grasp. The examples above are test-time executions of HANDFUL. • Pick Second: The robot must grasp and … view at source ↗
Figure 4
Figure 4. Figure 4: Performance of HANDFUL with curriculum learning vs without curriculum learning over full training with 10 million steps. test for our selection criteria: training is volatile due to tight constraints on in-hand space and finger placement, and this volatility is consequential. Accidentally eliminating the top 2-3 grasp candidates can cause a larger performance drop off compared to tasks such as Push Object,… view at source ↗
Figure 5
Figure 5. Figure 5: Left: the real world experiment setup with the xArm7 robot, LEAP hand, and the two Intel L515 cameras indicated with dotted boxes. Right: the different objects we use for experiments. Tasks Push Object Press Button Twist Knob Pull Drawer Pick Second Success (Both) 10 8 6 6 4 Fail Subtask 1 Only 4 4 6 6 7 Fail Subtask 2 Only 0 0 0 0 1 Fail Both 1 3 3 3 3 TABLE IV: Real-world experiment results. For each tas… view at source ↗
Figure 6
Figure 6. Figure 6: Hand poses in the human teleoperated data. Across tasks, operators predominantly relied on the index finger for the second subtask and used the remaining fingers for grasp stabilization, reflecting a consistent bias and limited pose diversity. VIII. LIMITATIONS Our experiments use a relatively controlled setting, with state-based observations and limited randomization (e.g., grasping only a block). This si… view at source ↗
Figure 7
Figure 7. Figure 7: Success rates of the nine grasping policies in simulation versus real-world settings. Each point represents a policy evaluated on either the black or red block. The dashed line y = x denotes the ideal case with no sim-to-real gap, where success rates in simulation and real world are identical. While in reality, simulation performance remains consistently high, real-world results exhibit substantial variabi… view at source ↗
read the original abstract

Dexterous robot hands offer rich opportunities for multifunctional manipulation, where a robot must execute multiple skills in sequence while maintaining control over previously grasped objects. Most prior work in dexterous manipulation focuses on single-object, single-skill tasks. In contrast, our insight is that many sequential tasks require resource-aware grasps that conserve fingers for future actions. In this paper, we study sequential grasp-conditioned dexterous manipulation, where a robot first grasps an object and then performs a second, distinct manipulation subtask while preserving the initial grasp. We introduce HANDFUL, a learning framework that models finger usage as a limited resource and encourages exploration of resource-aware grasps through finger-level contact rewards. These grasps are subsequently selected for downstream tasks via curriculum-based policy learning. We further propose HANDFUL-Bench, a simulation benchmark that introduces sequential dexterous manipulation tasks across multiple secondsubtask objectives, including pushing, pulling, and pressing, under a shared grasp-conditioned setup. Extensive simulation results demonstrate that prioritizing resource-aware grasps improves second-subtask success and robustness compared to a baseline that greedily optimizes the initial grasp before attempting the second subtask. We additionally validate our approach on a real dexterous LEAP hand. Together, this work establishes resource-aware grasp planning as a key principle for multifunctional dexterous manipulation. Supplementary material is available on our website: https://handful-dex.github.io.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper introduces HANDFUL, a reinforcement learning framework for sequential grasp-conditioned dexterous manipulation. It models finger usage as a limited resource and employs finger-level contact rewards to encourage grasps that leave fingers available for a subsequent distinct subtask (e.g., pushing, pulling, or pressing an object). The method uses curriculum-based policy learning to select such grasps, is evaluated on the new HANDFUL-Bench simulation benchmark across multiple second-subtask objectives, and is validated on a real LEAP hand. The central empirical claim is that resource-aware grasps improve second-subtask success and robustness relative to a greedy baseline that optimizes only the initial grasp.

Significance. If the results hold under matched conditions, the work would usefully highlight resource awareness as a design principle for multifunctional dexterous manipulation, moving beyond single-skill tasks. The HANDFUL-Bench benchmark and the real-robot demonstration are concrete contributions that could support follow-on research. The approach is entirely empirical (RL with hand-crafted rewards and curriculum), so it does not offer parameter-free derivations or machine-checked proofs, but the reproducible experimental setup on a public benchmark is a positive feature.

major comments (3)
  1. [Results / baseline description] Results section (baseline comparison): The abstract and methods describe the greedy baseline only at a high level as one that 'greedily optimizes the initial grasp.' It is not stated whether this baseline uses the identical policy architecture, action space, curriculum stage thresholds, and exploration noise schedule as HANDFUL. Without this matching, the reported gains in second-subtask success cannot be unambiguously attributed to the finger-level contact rewards rather than differences in the overall learning procedure.
  2. [Methods / reward and curriculum] Methods (reward formulation): The finger contact reward scales and curriculum stage thresholds are listed as free parameters. The central claim that these rewards produce grasps preserving both stability and finger availability rests on the empirical outcomes; however, no ablation is described that varies these scales while holding all other components fixed, leaving open whether the reported robustness is sensitive to hyperparameter choice.
  3. [Experiments / real-robot] Real-robot validation paragraph: The abstract states that the approach is validated on a physical LEAP hand, yet no quantitative details (number of trials, success metric definitions, or sim-to-real transfer protocol) are provided in the summary. This information is load-bearing for the robustness claim that extends beyond simulation.
minor comments (2)
  1. [Abstract] The abstract refers to 'secondsubtask' without a hyphen or space; consistent hyphenation should be used throughout.
  2. [Figures and tables] Figure captions and table headers should explicitly state the number of random seeds and total episodes used for each reported success rate to allow readers to assess statistical reliability.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We address each major comment point by point below, clarifying our approach and committing to revisions that strengthen the manuscript without altering its core claims.

read point-by-point responses

  1. Referee: [Results / baseline description] Results section (baseline comparison): The abstract and methods describe the greedy baseline only at a high level as one that 'greedily optimizes the initial grasp.' It is not stated whether this baseline uses the identical policy architecture, action space, curriculum stage thresholds, and exploration noise schedule as HANDFUL. Without this matching, the reported gains in second-subtask success cannot be unambiguously attributed to the finger-level contact rewards rather than differences in the overall learning procedure.

    Authors: We agree that explicit matching details are necessary to isolate the contribution of the finger-level contact rewards. In the revised manuscript, we will expand the Methods and Results sections to state that the greedy baseline employs exactly the same policy architecture, action space, curriculum stage thresholds, and exploration noise schedule as HANDFUL. The only difference is the absence of the finger contact reward term, so that performance differences can be attributed directly to resource-aware grasp selection. revision: yes

  2. Referee: [Methods / reward and curriculum] Methods (reward formulation): The finger contact reward scales and curriculum stage thresholds are listed as free parameters. The central claim that these rewards produce grasps preserving both stability and finger availability rests on the empirical outcomes; however, no ablation is described that varies these scales while holding all other components fixed, leaving open whether the reported robustness is sensitive to hyperparameter choice.

    Authors: We acknowledge that an ablation on the reward scales and curriculum thresholds would strengthen the robustness claim. We will add such an ablation study to the revised manuscript (or supplementary material), systematically varying these hyperparameters while keeping the policy architecture, curriculum structure, and other components fixed. This will show that the reported gains in second-subtask success and robustness hold across reasonable ranges of these parameters. revision: yes

  3. Referee: [Experiments / real-robot] Real-robot validation paragraph: The abstract states that the approach is validated on a physical LEAP hand, yet no quantitative details (number of trials, success metric definitions, or sim-to-real transfer protocol) are provided in the summary. This information is load-bearing for the robustness claim that extends beyond simulation.

    Authors: We will revise the real-robot validation section to include the requested quantitative details: the number of trials performed, precise definitions of the success metrics used, and a description of the sim-to-real transfer protocol (including any domain randomization or calibration steps). These additions will make the physical validation reproducible and directly support the robustness claims. revision: yes

Circularity Check

0 steps flagged

No significant circularity: empirical RL framework with independent validation

full rationale

The paper proposes HANDFUL as an RL-based framework using finger-level contact rewards to encourage resource-aware grasps, followed by curriculum policy learning and evaluation on HANDFUL-Bench. The central claims rest on simulation results comparing to a greedy baseline and real-robot validation, with no mathematical derivations, first-principles predictions, or fitted parameters presented as outputs. No self-citations, ansatzes, or renamings reduce the method to its inputs by construction; the approach is a standard empirical pipeline relying on external benchmarks and matched experimental conditions.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The approach depends on standard robotics simulation assumptions and RL training procedures, along with several tunable reward and curriculum parameters whose specific values are not detailed in the abstract.

free parameters (2)
  • Finger contact reward scales
    Weights balancing grasp stability against resource conservation in the RL objective.
  • Curriculum stage thresholds
    Parameters controlling progression between grasp learning and second-subtask policy training.
axioms (2)
  • domain assumption Simulation physics sufficiently approximates real dexterous hand dynamics for policy transfer
    Required for training in HANDFUL-Bench before real LEAP hand validation.
  • domain assumption RL can jointly optimize grasp and downstream manipulation under shared finger constraints
    Foundational to the curriculum-based policy learning approach.

pith-pipeline@v0.9.0 · 5571 in / 1440 out tokens · 70718 ms · 2026-05-07T16:14:58.056146+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

47 extracted references · 7 canonical work pages · 6 internal anchors

  1. [1]

    Dexart: Benchmarking generalizable dex- terous manipulation with articulated objects,

    C. Bao, H. Xu, Y . Qin, and X. Wang, “Dexart: Benchmarking generalizable dex- terous manipulation with articulated objects,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023

    2023

  2. [2]

    Sequential dexterity: Chaining dexterous policies for long-horizon manipulation,

    Y . Chen, C. Wang, L. Fei-Fei, and C. K. Liu, “Sequential dexterity: Chaining dexterous policies for long-horizon manipulation,” in Conference on Robot Learning (CoRL), 2023

    2023

  3. [3]

    PyBullet, a Python Module for Physics Simulation for Games, Robotics and Machine Learning,

    E. Coumans and Y . Bai, “PyBullet, a Python Module for Physics Simulation for Games, Robotics and Machine Learning,” 2016–2020

    2016

  4. [4]

    Mogrip: Gripper for multiobject grasping in pick-and-place tasks using translational movements of fingers,

    J. Eom, S. Y . Yu, W. Kim, C. Park, K. Y . Lee, and K.-J. Cho, “Mogrip: Gripper for multiobject grasping in pick-and-place tasks using translational movements of fingers,” Science Robotics, 2024

    2024

  5. [5]

    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 (ICML), 2018

    2018

  6. [6]

    Sequential multi-object grasping with one dexterous hand,

    S. He, Z. Shangguan, K. Wang, Y . Gu, Y . Fu, Y . Fu, and D. Seita, “Sequential multi-object grasping with one dexterous hand,” in IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2025

    2025

  7. [7]

    RLBench: The Robot Learning Benchmark & Learning Environment,

    S. James, Z. Ma, D. Rovick Arrojo, and A. J. Davison, “RLBench: The Robot Learning Benchmark & Learning Environment,” in IEEE Robotics and Automation Letters (RA-L), 2020

    2020

  8. [8]

    Learning to Singulate Objects in Packed Environments using a Dexterous Hand,

    H. Jiang, Y . Wang, H. Zhou, and D. Seita, “Learning to Singulate Objects in Packed Environments using a Dexterous Hand,” in International Symposium on Robotics Research (ISRR), 2024

    2024

  9. [9]

    Learning Geometry-Aware Non- prehensile Pushing and Pulling with Dexterous Hands,

    Y . Li, Y . Ling, G. S. Sukhatme, and D. Seita, “Learning Geometry-Aware Non- prehensile Pushing and Pulling with Dexterous Hands,” in IEEE International Conference on Robotics and Automation (ICRA), 2026

    2026

  10. [10]

    Grasp multiple objects with one hand,

    Y . Li, B. Liu, Y . Geng, P. Li, Y . Yang, Y . Zhu, T. Liu, and S. Huang, “Grasp multiple objects with one hand,” in IEEE Robotics and Automation Letters (RA-L), 2024

    2024

  11. [11]

    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,” in Conference on Robot Learning (CoRL), 2020

    2020

  12. [12]

    LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learning,

    B. Liu, Y . Zhu, C. Gao, Y . Feng, Q. Liu, Y . Zhu, and P. Stone, “LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learning,” in Neural Information Processing Systems, 2023

    2023

  13. [13]

    Deep differentiable grasp planner for high-dof grippers,

    M. Liu, Z. Pan, K. Xu, K. Ganguly, and D. Manocha, “Deep differentiable grasp planner for high-dof grippers,” in Robotics: Science and Systems (RSS), 2020

    2020

  14. [14]

    Synthesizing diverse and physically stable grasps with arbitrary hand structures using differentiable force closure estimator,

    T. Liu, Z. Liu, Z. Jiao, Y . Zhu, and S.-C. Zhu, “Synthesizing diverse and physically stable grasps with arbitrary hand structures using differentiable force closure estimator,” in IEEE Robotics and Automation Letters (RA-L), 2022

    2022

  15. [15]

    Grasping a handful: Sequential multi-object dexterous grasp generation,

    H. Lu, Y . Dong, Z. Weng, F. T. Pokorny, J. Lundell, and D. Kragic, “Grasping a handful: Sequential multi-object dexterous grasp generation,” in IEEE Robotics and Automation Letters (RA-L), 2025

    2025

  16. [16]

    Get a grip: Multi-finger grasp evaluation at scale enables robust sim-to-real transfer,

    T. G. W. Lum, A. H. Li, P. Culbertson, K. Srinivasan, A. D. Ames, M. Schwager, and J. Bohg, “Get a grip: Multi-finger grasp evaluation at scale enables robust sim-to-real transfer,” in Conference on Robot Learning (CoRL), 2024

    2024

  17. [17]

    Isaac Gym: High Performance GPU-Based Physics Simulation For Robot Learning

    V . Makoviychuk, L. Wawrzyniak, Y . Guo, M. Lu, K. Storey, M. Macklin, D. Hoeller, N. Rudin, A. Allshire, A. Handa, and G. State, “Isaac Gym: High Performance GPU-Based Physics Simulation For Robot Learning,” arXiv preprint arXiv:2108.10470, 2021

    work page internal anchor Pith review arXiv 2021

  18. [18]

    Isaac Lab: A GPU-Accelerated Simulation Framework for Multi-Modal Robot Learning

    M. Mittal, P. Roth, J. Tigue, A. Richard et al., “Isaac Lab: A GPU- Accelerated Simulation Framework for Multi-Modal Robot Learning,” arXiv preprint arXiv:2511.04831, 2025

    work page internal anchor Pith review arXiv 2025

  19. [19]

    RoboCasa: Large-Scale Simulation of Everyday Tasks for Generalist Robots,

    S. Nasiriany, A. Maddukuri, L. Zhang, A. Parikh, A. Lo, A. Joshi, A. Man- dlekar, and Y . Zhu, “RoboCasa: Large-Scale Simulation of Everyday Tasks for Generalist Robots,” in Robotics: Science and Systems (RSS), 2024

    2024

  20. [20]

    Solving Rubik's Cube with a Robot Hand

    OpenAI, I. Akkaya, M. Andrychowicz, M. Chociej, M. Litwin, B. McGrew, A. Petron, A. Paino, M. Plappert, G. Powell, R. Ribas, J. Schneider, N. Tezak, J. Tworek, P. Welinder, L. Weng, Q. Yuan, W. Zaremba, and L. Zhang, “Solving rubik’s cube with a robot hand,” arXiv preprint arXiv:1910.07113, 2019

    work page internal anchor Pith review arXiv 1910

  21. [21]

    In-Hand Object Rotation via Rapid Motor Adaptation,

    H. Qi, A. Kumar, R. Calandra, Y . Ma, and J. Malik, “In-Hand Object Rotation via Rapid Motor Adaptation,” in Conference on Robot Learning (CoRL), 2022

    2022

  22. [22]

    General In-Hand Object Rotation with Vision and Touch,

    H. Qi, B. Yi, S. Suresh, M. Lambeta, Y . Ma, R. Calandra, and J. Malik, “General In-Hand Object Rotation with Vision and Touch,” in Conference on Robot Learning (CoRL), 2023

    2023

  23. [23]

    DexPoint: Generalizable Point Cloud Reinforcement Learning for Sim-to-Real Dexterous Manipulation,

    Y . Qin, B. Huang, Z.-H. Yin, H. Su, and X. Wang, “DexPoint: Generalizable Point Cloud Reinforcement Learning for Sim-to-Real Dexterous Manipulation,” in Conference on Robot Learning (CoRL), 2022

    2022

  24. [24]

    SAM 2: Segment Anything in Images and Videos

    N. Ravi, V . Gabeur, Y .-T. Hu, R. Hu, C. Ryali, T. Ma, H. Khedr, R. R ¨adle, C. Rolland, L. Gustafson, E. Mintun, J. Pan, K. V . Alwala, N. Carion, C.-Y . Wu, R. Girshick, P. Doll ´ar, and C. Feichtenhofer, “SAM 2: Segment Anything in Images and Videos,” arXiv preprint arXiv:2408.00714, 2024

    work page internal anchor Pith review arXiv 2024

  25. [25]

    In-Hand Dexterous Manipulation of Piecewise-Smooth 3-D Objects,

    D. Rus, “In-Hand Dexterous Manipulation of Piecewise-Smooth 3-D Objects,” in International Journal of Robotics Research (IJRR), 1999

    1999

  26. [26]

    Proximal Policy Optimization Algorithms

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, “Proximal Policy Optimization Algorithms,” arXiv:1707.06347, 2017

    work page internal anchor Pith review arXiv 2017

  27. [27]

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

    D. Seita, P. Florence, J. Tompson, E. Coumans, V . Sindhwani, K. Goldberg, and A. Zeng, “Learning to Rearrange Deformable Cables, Fabrics, and Bags with Goal-Conditioned Transporter Networks,” in IEEE International Conference on Robotics and Automation (ICRA), 2021

    2021

  28. [28]

    Leap hand: Low-cost, efficient, and anthropomorphic hand for robot learning,

    K. Shaw, A. Agarwal, and D. Pathak, “Leap hand: Low-cost, efficient, and anthropomorphic hand for robot learning,” in Robotics: Science and Systems (RSS), 2023

    2023

  29. [29]

    Maniskill3: Gpu parallelized robotics simulation and rendering for generalizable embodied ai.arXiv preprint arXiv:2410.00425, 2024

    S. Tao, F. Xiang, A. Shukla, Y . Qin, X. Hinrichsen, X. Yuan, C. Bao, X. Lin, Y . Liu, T. kai Chan, Y . Gao, X. Li, T. Mu, N. Xiao, A. Gurha, Z. Huang, R. Calandra, R. Chen, S. Luo, and H. Su, “ManiSkill3: GPU Parallelized Robotics Simulation and Rendering for Generalizable Embodied AI,” arXiv preprint arXiv:2410.00425, 2024

    work page arXiv 2024

  30. [30]

    MuJoCo: A Physics Engine for Model- Based Control,

    E. Todorov, T. Erez, and Y . Tassa, “MuJoCo: A Physics Engine for Model- Based Control,” in IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2012

    2012

  31. [31]

    DexCap: Scal- able and Portable Mocap Data Collection System for Dexterous Manipulation,

    C. Wang, H. Shi, W. Wang, R. Zhang, L. Fei-Fei, and C. K. Liu, “DexCap: Scal- able and Portable Mocap Data Collection System for Dexterous Manipulation,” in Robotics: Science and Systems (RSS), 2024

    2024

  32. [32]

    Dexgraspnet: A large-scale robotic dexterous grasp dataset for general objects based on simulation,

    R. Wang, J. Zhang, J. Chen, Y . Xu, P. Li, T. Liu, and H. Wang, “Dexgraspnet: A large-scale robotic dexterous grasp dataset for general objects based on simulation,” in IEEE International Conference on Robotics and Automation (ICRA), 2023

    2023

  33. [33]

    SAPIEN: A SimulAted Part-based Interactive ENvironment,

    F. Xiang, Y . Qin, K. Mo, Y . Xia, H. Zhu, F. Liu, M. Liu, H. Jiang, Y . Yuan, H. Wang, L. Yi, A. X. Chang, L. J. Guibas, and H. Su, “SAPIEN: A SimulAted Part-based Interactive ENvironment,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020

    2020

  34. [34]

    Unidexgrasp: Universal robotic dexterous grasping via learning diverse proposal generation and goal-conditioned policy,

    Y . Xu, W. Wan, J. Zhang, H. Liu, Z. Shan, H. Shen, R. Wang, H. Geng, Y . Weng, J. Chen et al., “Unidexgrasp: Universal robotic dexterous grasping via learning diverse proposal generation and goal-conditioned policy,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023

    2023

  35. [35]

    Static grasp stability analysis of multiple spatial objects,

    T. Yamada and H. Yamamoto, “Static grasp stability analysis of multiple spatial objects,” Journal of Control Science and Engineering, vol. 3, 2015

    2015

  36. [36]

    Exploiting kinematic redundancy for robotic grasping of multiple objects,

    K. Yao and A. Billard, “Exploiting kinematic redundancy for robotic grasping of multiple objects,” in IEEE Transactions on Robotics, 2023

    2023

  37. [37]

    Rotating without Seeing: Towards In-hand Dexterity through Touch,

    Z.-H. Yin, B. Huang, Y . Qin, Q. Chen, and X. Wang, “Rotating without Seeing: Towards In-hand Dexterity through Touch,” in Robotics: Science and Systems (RSS), 2023

    2023

  38. [38]

    Optimization of power grasps for multiple objects,

    T. Yoshikawa, T. Watanabe, and M. Daito, “Optimization of power grasps for multiple objects,” inIEEE International Conference on Robotics and Automation (ICRA), 2001

    2001

  39. [39]

    Computation of grasp internal forces for stably grasping multiple objects,

    Y . Yu, K. Fukuda, and S. Tsujio, “Computation of grasp internal forces for stably grasping multiple objects,” in IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2001

    2001

  40. [40]

    MuJoCo Playground,

    K. Zakka, B. Tabanpour, Q. Liao, M. Haiderbhai, S. Holt, J. Y . Luo, A. Allshire, E. Frey, K. Sreenath, L. A. Kahrs, C. Sferrazza, Y . Tassa, and P. Abbeel, “MuJoCo Playground,” in Robotics: Science and Systems (RSS), 2025

    2025

  41. [41]

    MuJoCo Menagerie: A collection of high-quality simulation models for MuJoCo,

    K. Zakka, Y . Tassa, and MuJoCo Menagerie Contributors, “MuJoCo Menagerie: A collection of high-quality simulation models for MuJoCo,” 2022

    2022

  42. [42]

    3D Diffusion Policy: Generalizable Visuomotor Policy Learning via Simple 3D Representations,

    Y . Ze, G. Zhang, K. Zhang, C. Hu, M. Wang, and H. Xu, “3D Diffusion Policy: Generalizable Visuomotor Policy Learning via Simple 3D Representations,” in Robotics: Science and Systems (RSS), 2024

    2024

  43. [43]

    DexGraspNet 2.0: Learning Generative Dexterous Grasping in Large-scale Synthetic Cluttered Scenes,

    J. Zhang, H. Liu, D. Li, X. Yu, H. Geng, Y . Ding, J. Chen, and H. Wang, “DexGraspNet 2.0: Learning Generative Dexterous Grasping in Large-scale Synthetic Cluttered Scenes,” in Conference on Robot Learning (CoRL), 2024

    2024

  44. [44]

    robosuite: A Modular Simulation Framework and Benchmark for Robot Learning

    Y . Zhu, J. Wong, A. Mandlekar, R. Mart ´ın-Mart´ın, A. Joshi, S. Nasiriany, and Y . Zhu, “robosuite: A Modular Simulation Framework and Benchmark for Robot Learning,” in arXiv preprint arXiv:2009.12293, 2020. XI. APPENDIX A. Candidate Grasp Selection V alidation on All Tasks We extend the case study from Section VI-C.3 to all tasks. Table V and Table VI ...

    work page internal anchor Pith review arXiv 2009

  45. [45]

    Predictivity of Early Curriculum Stages:The pat- tern observed forPick Secondin the main paper holds broadly across tasks. The top-performing candidates in the final curriculum stage consistently rank among the leaders in earlier stages, confirming that early performance is a reliable signal for grasp selection. Early curriculum stages seem more predictiv...

  46. [46]

    The best-performing candidate grasp is never prematurely eliminated in any seed of any task, and the second and third-best candidates are often retained

    Selection Stability Across Seeds:Selection stability is similarly strong across tasks. The best-performing candidate grasp is never prematurely eliminated in any seed of any task, and the second and third-best candidates are often retained. As inPick Second, there is some volatility in selecting these second and third-best grasps, but the impact is minor ...

  47. [47]

    In our experi- ments, candidates such asπ 1 andπ 8 perform strongly across nearly all tasks, which may reflect broad task compatibility, high grasp stability in this seed, or both

    Limitations:Because we train only a single grasping seed per candidate strategy, it is difficult to cleanly attribute each grasp’s second-subtask performance to its intrinsic task suitability versus the stability of that particular grasp instance (which may vary with the training seed). In our experi- ments, candidates such asπ 1 andπ 8 perform strongly a...