Pith. sign in

REVIEW 3 major objections 5 minor 2 cited by

This paper claims that dexterous hands can keep tools firmly gripped during hammering, sawing, cutting, stirring, and scooping by combining wrench-tested grasp selection with a reinforcement-learning controller that issues small finger corr

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 21:19 UTC pith:S63UJMEU

load-bearing objection The core integration is new and the hardware results are encouraging, but the real-world evidence is underpowered and the 'consistently outperforming' claim overreaches. the 3 major comments →

arxiv 2602.20466 v1 pith:S63UJMEU submitted 2026-02-24 cs.RO

Grasp to Act: Dexterous Grasping for Tool Use in Dynamic Settings

Harsh Gupta , Mohammad Amin Mirzaee , Wenzhen Yuan This is my paper
classification cs.RO
keywords dexterous graspingtool usewrench-space stabilitysim-to-realreinforcement learninggrasp optimizationin-hand slipdynamic manipulation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tries to establish that dexterous robotic hands can hold tools securely while the tools are doing real physical work, not just while being lifted. Its recipe is two-stage: first pick a grasp by testing candidate hand configurations against strong forces and torques from many directions in simulation, then run a reinforcement-learned policy that makes small finger-joint corrections during the motion to cancel in-hand slip. The authors report that this combination transfers from simulation to a real hand with zero fine-tuning and beats both pure grasp-optimization and pure RL baselines, reducing slip and raising task-completion rates. If true, it would make grasping a stepping stone to functional manipulation rather than a static goal.

Core claim

Grasp-to-Act claims that stable tool use emerges from two complementary steps rather than either alone. The first offline step samples many candidate grasps around the region of a human demonstration, then scores each by applying linearly increasing external wrenches in all twelve force/torque directions in simulation and measuring how long the object stays within slip thresholds; the top-scoring enveloping grasps survive. The second online step is a residual RL controller that starts from that grasp and, at 30 Hz, reads only joint positions, joint targets, and joint torques plus the target object trajectory, and outputs small joint corrections to keep the tool aligned. In simulation and in

What carries the argument

The central mechanism is the pairing of a wrench-space stability score, which ranks candidate grasps by their average survival time under 12 directions of external force and torque, with a residual reinforcement-learning policy that maps proprioceptive signals — joint positions, joint targets, and joint torques — to small finger-joint adjustments, never observing the object’s pose directly. The score selects grasps with broad stability regions; the residual policy keeps the grasp inside that region when real disturbances push the object off target.

Load-bearing premise

The simplified simulation forces — a linear resistance term and a constant directional application force, calibrated once per material — match real disturbances closely enough that a policy trained on them, using only joint signals, can correct every slip that occurs on hardware.

What would settle it

Run the same five tasks on a real 16-degree-of-freedom dexterous hand with the same synthesized grasps but introduce a disturbance outside the calibrated force family — for instance, a hammer strike at an off-axis angle or a saw blade that catches intermittently — and measure whether in-hand translational and rotational slip accumulate beyond the reported thresholds until the tool drops; repeat across many trials to see whether the gap between simulation-trained expectations and real slip grows.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Static lift-based grasp metrics are insufficient for functional tool use; disturbance-wrench tests along all force and torque axes give a more predictive stability ranking.
  • Initializing RL from an analytically stable grasp narrows the exploration space, so the policy converges quickly and focuses on slip correction rather than learning a grasp from scratch.
  • Zero-shot sim-to-real for contact-rich tool use is feasible without modeling fluids or deformables in detail, using calibrated proxy forces during training.
  • The policy’s reliance on proprioception only means the hand does not need external object tracking during execution, simplifying deployment on physical robots.
  • The method generalizes to unseen variants of a tool (e.g., different hammer masses and handle diameters) without retraining, at least within the reported test range.

Where Pith is reading between the lines

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

  • If the hardware results extend beyond the 10 trials reported per task, the same hybrid recipe could apply to other dexterous tasks involving sustained loads, such as drilling, polishing, or screwdriving, with task-specific force calibration.
  • Because the policy detects slip only through joint-torque signatures, its ceiling depends on every failure mode producing a detectable proprioceptive pattern; adding tactile sensing or an explicit slip observer would likely broaden the range of correctable slips.
  • The real-world comparison would be strengthened by reporting variance or significance across trials; the current figure, shown as 10-trial averages without error bars, is suggestive rather than definitive about the margin over baselines.
  • A direct stress test would be to run the same grasps with disturbances outside the calibrated force family, such as off-axis hammer impacts or intermittent saw catch, to expose whether the simplified simulation forces are the limiting assumption.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes Grasp-to-Act (G2A), a two-stage pipeline for dexterous tool use. Stage 1 samples candidate grasps around a human-demonstrated wrist pose and ranks them by wrench-space stability tests performed in Isaac Lab. Stage 2 trains a PPO residual policy that adjusts finger joint targets from proprioceptive observations, task phase, and the demonstrated target trajectory, under simplified force models for resistive and application forces. The method is evaluated in simulation across five tasks (hammering, sawing, cutting, stirring, scooping) against six baselines, and on a real LEAP hand/UR5 platform with 10 trials per task, reporting in-hand slip and task-completion metrics. The central claim is that G2A enables robust zero-shot sim-to-real transfer and consistently outperforms baselines on both slip and task completion.

Significance. The direction is timely and the hybrid design is well motivated: analytical wrench-space grasp scoring is used to initialize a residual RL controller, rather than relying on either component alone. The use of human demonstrations only to define the functional grasp region, while still allowing the hand to discover its own morphology-specific grasps, is a sensible way to sidestep direct retargeting. The hardware results, if statistically reliable, would be a useful demonstration that static grasp metrics are insufficient for dynamic tool use and that online fingertip adaptation matters. The paper also makes an effort to calibrate force coefficients from physical measurements and to domain-randomize during training. However, the evidence for the paper's strongest claim is limited by the simplified sim environment and, more importantly, by the underpowered real-world evaluation.

major comments (3)
  1. [Section V-B, Fig. 5] The central real-world claim of consistent outperformance rests on 10 trials per task with no error bars, confidence intervals, or significance tests. In tasks where G2A w/o Adaptation is already near ceiling (stirring and scooping), a single trial difference changes the reported completion by 10 points, so the ordering may not be robust. Please report per-trial data, add paired tests appropriate for small samples (e.g., Wilcoxon signed-rank or exact binomial tests), and give confidence intervals or effect sizes. Without this, the statement that G2A 'consistently outperforms' baselines is not established.
  2. [Section III-C.1, Section III-C.2] The simulation replaces the actual task physics with F_res = k_res * v_eff and F_app = k_app * d_hat, and the policy never observes the object pose, relying instead on proprioceptive signatures. The zero-shot transfer claim therefore depends on (i) the adequacy of these linear/constant force models and (ii) the informativeness of joint positions/targets/torques for detecting and correcting slip. Table I validates only the simplified simulation, not the real contact-rich physics. Please add a quantitative comparison between the modeled wrench profiles and measured real-world disturbance forces, or an ablation where the policy observes object pose, or an analysis of failure cases showing that proprioception alone is sufficient. This is needed to support the sim-to-real claim.
  3. [Section IV-D, Table I] The simulation metric S is defined only as 'object remains held without dropping'; it does not measure task completion for hammering, sawing, cutting, stirring, or scooping because these interactions are not explicitly modeled. The paper acknowledges this, but it means Table I cannot support claims about functional task success. The real-world task-completion metrics in Section IV-B are the only evidence for the actual tasks, and they are the same underpowered 10-trial measurements. Please either add task-level success metrics in simulation or clearly state, in the abstract and results, that simulation validates only grasp retention under idealized force profiles.
minor comments (5)
  1. [Section II.A] The heading uses 'Dextrous Grasp Synthesis'; elsewhere the paper uses 'dexterous.' Please standardize the spelling.
  2. [Section II.A, references] Reference [17] is a contrastive-learning paper (Robinson et al., ICLR 2021) and appears mismatched for the claim about continuous stability scores. Please check and replace with the intended grasp-ranking reference.
  3. [Table II] The rows for the three hammer variants are not labeled beyond the first 'Hammer' row. Adding explicit variant labels (e.g., mass/handle diameter) would make the generalization result reproducible.
  4. [Section III-B.2] The stability scoring depends on thresholds (Fmax, tau_max, tmax, delta_p, delta_theta) but their numerical values are not reported in the main text. Please include them, or state that they are in the appendix/website.
  5. [Section VI] The discussion correctly notes that the initial grasp region must be unobstructed and that the sawing task requires manual staging of the handle loop. This limitation should be reflected in the abstract or conclusion, where the method is described as fully autonomous across the five tasks.

Circularity Check

0 steps flagged

No significant circularity: the core sim-to-real claim is evaluated by external task metrics and calibrated force models, not by reusing fitted parameters or self-citations.

full rationale

The paper's derivation chain is not circular in any load-bearing sense. The initial grasp ranking uses wrench-space stability tests with position/orientation thresholds δp/δθ (Sec. III-B.2), and the RL residual controller receives a penalty when the same thresholds are exceeded (Sec. III-C.2). However, the thresholds are consistency constants, not fitted outputs, and the central real-world claims are measured with external metrics: task completion T is defined functionally for each task (e.g., nail embedding fraction, sawed length, transferred mass, Sec. IV-B) and slip is reported as continuous distance/angle errors E_t and E_θ from OptiTrack tracking (Sec. IV-D). Training a policy to minimize trajectory-tracking error and then measuring trajectory-tracking error is standard objective/evaluation alignment, not a prediction forced by construction. The simplified simulation forces F_res = k_res·v_eff and F_app = k_app·d̂ (Sec. III-C.1) are calibrated from physical measurements (force gauge, measured masses), not fit to the evaluated outcomes, so the real-world results are not statistically forced by the training inputs. There is no self-citation chain that carries the central argument, no imported uniqueness theorem, and no ansatz presented as derived from cited prior work. The paper's own Discussion identifies the main limitations (unobstructed grasp regions, manual staging for sawing, simplified geometries, task-specific policies), which are honest scope restrictions rather than hidden circularity. The remaining concern about 10 trials per task without significance tests is an evidence-quality issue, not a circularity issue.

Axiom & Free-Parameter Ledger

10 free parameters · 5 axioms · 0 invented entities

The ledger shows the paper buys many things from the environment: pose estimators, simulator physics, and task-specific force abstractions. Most free parameters are hand-chosen sampling/threshold values that are not reported; the two force coefficients are calibrated from direct measurements. No new physical entities are invented. The most fragile item is the force abstraction + proprioception-only observation pairing, which bears most of the zero-shot transfer load.

free parameters (10)
  • Wrist perturbation limits (Δx, Δy, Δz, Δθx, Δθy, Δθz)
    Sampling box around the human wrist pose (§III-B.1.a) defines which grasps are even considered; bounds are not reported in the text.
  • Inter-finger angle sampling ranges (θf, θt)
    Controls finger spread during grasp initiation (§III-B.1.b); ranges not reported.
  • Joint-group closing rate ranges (q̇g)
    Independent closing speeds per finger joint group produce claw-like vs power grasps (§III-B.1.b); ranges not reported.
  • Joint torque threshold distribution (τq*)
    Fingers stop closing when torque exceeds a uniformly sampled threshold (§III-B.1.b); distribution/range omitted.
  • Wrench-test parameters (Fmax, τmax, tmax)
    Maximum disturbance magnitude and duration used in stability scoring (§III-B.2); values omitted.
  • Slip thresholds (δp, δθ)
    Define 'slip' in wrench tests and trigger the large penalty in RL (§III-B.2, III-C.2); values omitted and reused across scoring and metrics.
  • Resistive force coefficient (k_res) = calibrated with a force gauge at task speeds
    F_res=k_res·v_eff; calibrated per material, not fit to task success, but exact values/calibration protocol are not reported.
  • Application force coefficient (k_app) = set from measured payload mass
    F_app=k_app·d̂; used for payload/normal forces; exact values not reported.
  • Reward coefficients (αp, αq) and penalty magnitude (r_pen)
    Reward shaping in §III-C.2; values not reported, making the policy hard to re-train exactly.
  • Domain-randomization ranges = ±30% for mass, friction, force coefficients
    Nominal values being randomized are not stated; the ±30% range itself is a hand choice (§III-C.1).
axioms (5)
  • domain assumption FoundationPose and HaMeR-plus-ICP provide accurate 6D object and hand pose estimates from RGB-D video.
    The demo trajectory, grasp region, and real-world object pose chain all depend on these off-the-shelf estimators (§III-A, §IV-A); no error analysis is given.
  • domain assumption Isaac Lab's rigid-contact simulation is faithful enough to select stable grasps and train policies that transfer.
    All grasp scoring and RL training happen in Isaac Lab (§III-B/C); inaccurate contacts would invalidate both the ranking and the learned policy.
  • ad hoc to paper Task interactions with fluids and deformables reduce to F_res=k_res·v_eff and F_app=k_app·d̂.
    Introduced specifically to avoid simulating batter, cucumber, drywall, and beans; the paper validates coefficients only by calibration, not by comparing predicted vs measured force profiles during real tasks (§III-C.1).
  • domain assumption Proprioceptive observations (joint positions, targets, torques) are sufficient to detect and correct in-hand slip.
    Object pose is deliberately excluded from the policy observation (§III-C.2); the policy can only adapt if slip changes proprioceptive signals.
  • domain assumption The human-demonstrated object trajectory is a valid open-loop wrist reference with a fixed grasp offset.
    Wrist target is computed as T_hand^t=T_obj^t·T_grasp (§III-C); if the object shifts in hand, this reference becomes inconsistent and the wrist cannot compensate because its path is open-loop.

pith-pipeline@v1.3.0-alltime-deepseek · 11973 in / 17058 out tokens · 164868 ms · 2026-08-02T21:19:54.330538+00:00 · methodology

0 comments
read the original abstract

Achieving robust grasping with dexterous hands remains challenging, especially when manipulation involves dynamic forces such as impacts, torques, and continuous resistance--situations common in real-world tool use. Existing methods largely optimize grasps for static geometric stability and often fail once external forces arise during manipulation. We present Grasp-to-Act, a hybrid system that combines physics-based grasp optimization with reinforcement-learning-based grasp adaptation to maintain stable grasps throughout functional manipulation tasks. Our method synthesizes robust grasp configurations informed by human demonstrations and employs an adaptive controller that residually issues joint corrections to prevent in-hand slip while tracking the object trajectory. Grasp-to-Act enables robust zero-shot sim-to-real transfer across five dynamic tool-use tasks--hammering, sawing, cutting, stirring, and scooping--consistently outperforming baselines. Across simulation and real-world hardware trials with a 16-DoF dexterous hand, our method reduces translational and rotational in-hand slip and achieves the highest task completion rates, demonstrating stable functional grasps under dynamic, contact-rich conditions.

Figures

Figures reproduced from arXiv: 2602.20466 by Harsh Gupta, Mohammad Amin Mirzaee, Wenzhen Yuan.

Figure 1
Figure 1. Figure 1: Overview of the Grasp-to-Act framework. (Left) A human demonstration trajectory provides the reference of the dynamic conditions during tool using, as well as the initial grasping locations. Accordingly, we sample a series of grasps and score them to identify stable grasp candidates (middle left). A reinforcement learning policy then performs online-grasp adaptation by adjusting finger joint positions unde… view at source ↗
Figure 2
Figure 2. Figure 2: Pipeline of sampling and evaluating grasps. (A) We initialize the range of grasp locations based on human demonstration. (B-C) [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Real-world experiment setup. The workspace includes an (1) AprilTag, (2) tool tracking markers, (3) hand tracking markers, and an (4) RGBD camera. joint configuration. Formally, qt+1 = qt + ∆qt, where qt+1 are the executed joint positions. The reward function combines separate terms for position alignment, orientation alignment, and penalties for large deviations. We define the position reward as using a m… view at source ↗
Figure 4
Figure 4. Figure 4: Human demonstrations of the five functional tasks we [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Real-world results across five functional tasks. Comparison of (a) In-hand translational slip distance Et (cm), (b) In-hand slip rotation distance Eθ ( ◦ ), and (c) task completion T . Task-specific definitions of T are provided in Section IV-B. on the thinner-handled hammer. Our full G2A method sam￾ples grasps stable under large forces and torques—a broad stability region—and adapts online, achieving 100%… view at source ↗
Figure 6
Figure 6. Figure 6: Real-world rollouts. In-hand slip rotation distance Eθ for one representative rollout per baseline across the tasks. Task snapshots show the G2A execution. In practice, objects rest on tables or shelves, where contact with the surface can obstruct part of the graspable area; for sawing, we manually stage the tool with the handle loop ac￾cessible. In such cases, our framework cannot autonomously execute the… view at source ↗

discussion (0)

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

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Function-based Parametric Co-Design Optimization of Dexterous Hands

    cs.RO 2026-04 unverdicted novelty 6.0

    A unified parametric framework optimizes dexterous hand designs by combining structure, kinematics, and fine surface geometry for grasp stability in simulation and real-world tests.

  2. LUCID: Learning Embodiment-Agnostic Intent Models from Unstructured Human Videos for Scalable Dexterous Robot Skill Acquisition

    cs.RO 2026-06 unverdicted novelty 5.0

    LUCID learns embodiment-agnostic intent models from unstructured human videos to train dexterous robot policies in simulation, enabling zero-shot transfer on real-world tasks like stirring and wiping.

Reference graph

Works this paper leans on

37 extracted references · 13 linked inside Pith · cited by 2 Pith papers

  1. [1]

    Data-driven grasp synthesis—a survey,

    J. Bohg, A. Morales, T. Asfour, and D. Kragic, “Data-driven grasp synthesis—a survey,”IEEE Transactions on robotics, vol. 30, no. 2, pp. 289–309, 2013

    2013

  2. [2]

    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,” in2023 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2023, pp. 11 359–11 366

    2023

  3. [3]

    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,”arXiv preprint arXiv:2410.23701, 2024

    Pith/arXiv arXiv 2024

  4. [4]

    Web2grasp: Learning functional grasps from web images of hand-object interactions,

    H. Chen, Y . Yao, Y . Ye, Z. Xu, H. Bharadhwaj, J. Wang, S. Tulsiani, Z. Erickson, and J. Ichnowski, “Web2grasp: Learning functional grasps from web images of hand-object interactions,”arXiv preprint arXiv:2505.05517, 2025

    Pith/arXiv arXiv 2025

  5. [5]

    Graspit! a versatile simulator for robotic grasping,

    A. T. Miller and P. K. Allen, “Graspit! a versatile simulator for robotic grasping,”IEEE Robotics & Automation Magazine, vol. 11, no. 4, pp. 110–122, 2004

    2004

  6. [6]

    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,”IEEE Robotics and Automation Letters, vol. 7, no. 1, pp. 470–477, 2021

    2021

  7. [7]

    Dexterous grasping via eigengrasps: A low-dimensional approach to a high-complexity prob- lem,

    M. Ciocarlie, C. Goldfeder, and P. Allen, “Dexterous grasping via eigengrasps: A low-dimensional approach to a high-complexity prob- lem,” inRobotics: Science and systems manipulation workshop- sensing and adapting to the real world, 2007

    2007

  8. [8]

    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,”Robotics: Science and Systems (RSS), 2023

    2023

  9. [9]

    Grasp synthesis in cluttered en- vironments for dexterous hands,

    D. Berenson and S. S. Srinivasa, “Grasp synthesis in cluttered en- vironments for dexterous hands,” inHumanoids 2008-8th IEEE-RAS International Conference on Humanoid Robots. IEEE, 2008, pp. 189–196

    2008

  10. [10]

    Learning diverse and physically feasi- ble dexterous grasps with generative model and bilevel optimization,

    A. Wu, M. Guo, and C. K. Liu, “Learning diverse and physically feasi- ble dexterous grasps with generative model and bilevel optimization,” arXiv preprint arXiv:2207.00195, 2022

    Pith/arXiv arXiv 2022

  11. [11]

    Frogger: Fast robust grasp generation via the min-weight metric,

    A. H. Li, P. Culbertson, J. W. Burdick, and A. D. Ames, “Frogger: Fast robust grasp generation via the min-weight metric,” in2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2023, pp. 6809–6816

    2023

  12. [12]

    Planning optimal grasps,

    C. Ferrari, J. Cannyet al., “Planning optimal grasps,” inProceedings., 1992 IEEE International Conference on Robotics and Automation, 1992., vol. 3. IEEE, 1992, pp. 2290–2295

    1992

  13. [13]

    Dexonomy: Synthesizing all dexterous grasp types in a grasp taxonomy,

    J. Chen, Y . Ke, L. Peng, and H. Wang, “Dexonomy: Synthesizing all dexterous grasp types in a grasp taxonomy,”Robotics: Science and Systems, 2025

    2025

  14. [14]

    The grasp taxonomy of human grasp types,

    T. Feix, J. Romero, H.-B. Schmiedmayer, A. M. Dollar, and D. Kragic, “The grasp taxonomy of human grasp types,”IEEE Transactions on human-machine systems, vol. 46, no. 1, pp. 66–77, 2015

    2015

  15. [15]

    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, T. Liu, L. Yi, and H. Wang, “Unidexgrasp: Universal robotic dexterous grasping via learning diverse proposal generation and goal-conditioned policy,” inProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (CVPR), June 2023, pp. 4737–4746

    2023

  16. [16]

    Gendexgrasp: Generalizable dexterous grasping,

    P. Li, T. Liu, Y . Li, Y . Zhu, Y . Yang, and S. Huang, “Gendexgrasp: Generalizable dexterous grasping,”arXiv preprint arXiv:2210.00722, 2022

    Pith/arXiv arXiv 2022

  17. [17]

    Contrastive learning with hard negative samples,

    J. D. Robinson, C.-Y . Chuang, S. Sra, and S. Jegelka, “Contrastive learning with hard negative samples,” inInternational Conference on Learning Representations, 2021

    2021

  18. [18]

    Multi- grippergrasp: A dataset for robotic grasping from parallel jaw grippers to dexterous hands,

    L. F. Casas, N. Khargonkar, B. Prabhakaran, and Y . Xiang, “Multi- grippergrasp: A dataset for robotic grasping from parallel jaw grippers to dexterous hands,” in2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2024, pp. 2978–2984

    2024

  19. [19]

    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,” inConference on Robot Learning. PMLR, 2023, pp. 2549–2564

    2023

  20. [20]

    Dextrah- rgb: Visuomotor policies to grasp anything with dexterous hands,

    R. Singh, A. Allshire, A. Handa, N. Ratliff, and K. Van Wyk, “Dextrah- rgb: Visuomotor policies to grasp anything with dexterous hands,” arXiv preprint arXiv:2412.01791, 2024

    Pith/arXiv arXiv 2024

  21. [21]

    Visual dexterity: In-hand reorientation of novel and complex object shapes,

    T. Chen, M. Tippur, S. Wu, V . Kumar, E. Adelson, and P. Agrawal, “Visual dexterity: In-hand reorientation of novel and complex object shapes,”Science Robotics, vol. 8, no. 84, p. eadc9244, 2023

    2023

  22. [22]

    Learning complex dexterous manipulation with deep reinforcement learning and demonstrations,

    A. Rajeswaran, V . Kumar, A. Gupta, G. Vezzani, J. Schulman, E. Todorov, and S. Levine, “Learning complex dexterous manipulation with deep reinforcement learning and demonstrations,”arXiv preprint arXiv:1709.10087, 2017

    Pith/arXiv arXiv 2017

  23. [23]

    Crossing the human-robot embodiment gap with sim-to-real rl using one human demonstration,

    T. G. W. Lum, O. Y . Lee, C. K. Liu, and J. Bohg, “Crossing the human-robot embodiment gap with sim-to-real rl using one human demonstration,”arXiv preprint arXiv:2504.12609, 2025

    Pith/arXiv arXiv 2025

  24. [24]

    Fungrasp: Functional grasping for diverse dexterous hands,

    L. Huang, H. Zhang, Z. Wu, S. Christen, and J. Song, “Fungrasp: Functional grasping for diverse dexterous hands,”IEEE Robotics and Automation Letters, 2025

    2025

  25. [25]

    Dexterous functional grasping,

    A. Agarwal, S. Uppal, K. Shaw, and D. Pathak, “Dexterous functional grasping,”arXiv preprint arXiv:2312.02975, 2023

    Pith/arXiv arXiv 2023

  26. [26]

    Dextrah-g: Pixels-to- action dexterous arm-hand grasping with geometric fabrics,

    T. G. W. Lum, M. Matak, V . Makoviychuk, A. Handa, A. Allshire, T. Hermans, N. D. Ratliff, and K. Van Wyk, “Dextrah-g: Pixels-to- action dexterous arm-hand grasping with geometric fabrics,”arXiv preprint arXiv:2407.02274, 2024

    Pith/arXiv arXiv 2024

  27. [27]

    Lessons from learning to spin

    J. Wang, Y . Yuan, H. Che, H. Qi, Y . Ma, J. Malik, and X. Wang, “Lessons from learning to spin” pens”,”arXiv preprint arXiv:2407.18902, 2024

    Pith/arXiv arXiv 2024

  28. [28]

    Crossing the reality gap: A survey on sim-to-real transferability of robot controllers in reinforcement learning,

    E. Salvato, G. Fenu, E. Medvet, and F. A. Pellegrino, “Crossing the reality gap: A survey on sim-to-real transferability of robot controllers in reinforcement learning,”IEEE Access, vol. 9, pp. 153 171–153 187, 2021

    2021

  29. [29]

    Foundationpose: Unified 6d pose estimation and tracking of novel objects,

    B. Wen, W. Yang, J. Kautz, and S. Birchfield, “Foundationpose: Unified 6d pose estimation and tracking of novel objects,” inProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 17 868–17 879

    2024

  30. [30]

    Grounded sam: Assembling open-world models for diverse visual tasks,

    T. Ren, S. Liu, A. Zeng, J. Lin, K. Li, H. Cao, J. Chen, X. Huang, Y . Chen, F. Yanet al., “Grounded sam: Assembling open-world models for diverse visual tasks,”arXiv preprint arXiv:2401.14159, 2024

    Pith/arXiv arXiv 2024

  31. [31]

    Reconstructing hands in 3d with transformers,

    G. Pavlakos, D. Shan, I. Radosavovic, A. Kanazawa, D. Fouhey, and J. Malik, “Reconstructing hands in 3d with transformers,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 9826–9836

    2024

  32. [32]

    Embodied hands: Mod- eling and capturing hands and bodies together,

    J. Romero, D. Tzionas, and M. J. Black, “Embodied hands: Mod- eling and capturing hands and bodies together,”arXiv preprint arXiv:2201.02610, 2022

    Pith/arXiv arXiv 2022

  33. [33]

    Orbit: A unified simulation framework for interactive robot learning environments,

    M. Mittal, C. Yu, Q. Yu, J. Liu, N. Rudin, D. Hoeller, J. L. Yuan, R. Singh, Y . Guo, H. Mazhar, A. Mandlekar, B. Babich, G. State, M. Hutter, and A. Garg, “Orbit: A unified simulation framework for interactive robot learning environments,”IEEE Robotics and Automa- tion Letters, vol. 8, no. 6, pp. 3740–3747, 2023

    2023

  34. [34]

    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

    Pith/arXiv arXiv 2017

  35. [35]

    Long short-term memory,

    S. Hochreiter and J. Schmidhuber, “Long short-term memory,”Neural computation, vol. 9, no. 8, pp. 1735–1780, 1997

    1997

  36. [36]

    Apriltag: A robust and flexible visual fiducial system,

    E. Olson, “Apriltag: A robust and flexible visual fiducial system,” in2011 IEEE international conference on robotics and automation. IEEE, 2011, pp. 3400–3407

    2011

  37. [37]

    Dexsingrasp: Learning a unified policy for dexterous object singulation and grasping in cluttered environments,

    L. Xu, Z. Liu, Z. Gui, J. Guo, Z. Jiang, Z. Xu, C. Gao, and L. Shao, “Dexsingrasp: Learning a unified policy for dexterous object singulation and grasping in cluttered environments,”arXiv preprint arXiv:2504.04516, 2025

    arXiv 2025