Pith. sign in

REVIEW 1 major objections 8 minor 38 references

One teleoperation pipeline, three robots, eight objects

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 · glm-5.2

2026-07-08 21:31 UTC pith:HSQR5STQ

load-bearing objection Functional teleoperation system with a load-bearing gap in the grasping module's optimization claim the 1 major comments →

arxiv 2607.05883 v1 pith:HSQR5STQ submitted 2026-07-07 cs.RO

DexTele: A Dual-Arm Dexterous Teleoperation System Based on Motion Retargeting and Adaptive Force Control

Yuanchuan Lai , Qing Gao , Ziyan Liang , Xianfeng Cheng , Junjie Hu , Zhaojie Ju This is my paper
classification cs.RO
keywords teleoperationmotion retargetinggraph neural networkadaptive graspingforce controlvision-language modelmodel predictive controldual-arm 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.

DexTele is a dual-arm teleoperation system that maps human motion to multiple robot platforms and grasps diverse objects with adaptive force control. The system has two core mechanisms. The first is a graph neural network (SAG-GCN) with a dual-stream architecture that encodes human arm and hand motions as graph structures, optimizes them in a latent space, and decodes them into robot joint angles—without requiring paired human-robot training data for each new platform. The second is an adaptive grasping module where a vision-language model infers a target grasping force for a seen object (e.g., 300g for a water bottle), and a model predictive control loop uses a trained joint-angle-to-force surrogate model to iteratively adjust finger commands until the applied force converges to the target. Experiments show the retargeting module achieves lower position, rotation, and velocity errors than three baselines across two datasets and generalizes to three robot platforms (RMC-DA, YuMi, Unitree H1). The grasping module successfully grasps eight object categories with force deviations and oscillations under 10%, improving average grasp success from roughly 5/10 to 9/10 compared to operating without the adaptive module.

Core claim

The paper demonstrates that cross-platform motion retargeting can be treated as a graph-based latent optimization problem solved by a dual-stream graph convolutional network, eliminating the need for platform-specific paired training data. Separately, it shows that combining a vision-language model's semantic force estimates with gradient-based online optimization of a learned force surrogate yields compliant grasping across objects of varying rigidity and fragility. The two modules together form a complete teleoperation pipeline: vision-based human motion capture feeds the retargeting network, whose hand output is refined by the adaptive force controller before execution on the robot.

What carries the argument

SAG-GCN (Spatial Attention Gated Graph Convolutional Network) with dual-stream encoder-decoder; VLM-inferred target force; random-forest joint-angle-to-force surrogate; MPC-inspired gradient descent optimization loop

Load-bearing premise

The adaptive grasping module depends on a vision-language model to output appropriate target grasping forces for objects it sees (e.g., 300g for a water bottle, 30g for a paper cup), but the paper does not describe how these force values were determined, validated, or whether they generalize to objects of the same category with different masses, materials, or fragility.

What would settle it

If the SAG-GCN retargeting produces errors comparable to or worse than baseline methods on a fourth robot platform not used during development, or if the VLM-inferred target forces lead to object damage or grasp failure on a substantial fraction of novel object instances within stated categories.

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

If this is right

  • If the graph-based retargeting approach generalizes as claimed, adding a new robot platform requires only its URDF kinematic description, not new paired datasets or retraining from scratch.
  • The VLM-plus-MPC grasping architecture could be extended to any object the VLM can recognize, potentially scaling adaptive grasping to long-tail object categories without per-object engineering.
  • The 10 FPS real-time performance suggests the pipeline is fast enough for interactive teleoperation, though the authors note pose estimation latency is the current bottleneck.
  • The dual-stream separation of arm and hand motions could inform other domains where coarse and fine motor control operate at different scales, such as whole-body humanoid teleoperation.

Where Pith is reading between the lines

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

  • The VLM's force estimates are treated as a black box; if the VLM misjudges an object's fragility or mass (e.g., a thin-walled glass vs. a sturdy plastic bottle of the same category), the MPC will optimize toward an incorrect target force. The paper does not analyze failure modes of this pipeline.
  • The cross-platform retargeting is validated on robots with similar upper-body kinematics (5-7 DOF arms, 6-12 DOF hands). Whether the graph-based approach handles radically different morphologies (e.g., continuum robots, highly underactuated hands) remains untested.
  • The random forest force surrogate is trained on historical data from a specific dexterous hand; transferring the adaptive grasping module to a different hand would require retraining this surrogate, which is not discussed.
  • The grasping evaluation uses 10 trials per object; the reported success rates (e.g., 9/10 vs. 4/10) have wide confidence intervals, making fine-grained comparisons between object categories statistically weak.

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

1 major / 8 minor

Summary. This paper presents DexTele, a dual-arm dexterous teleoperation system combining a vision-based motion retargeting module (SAG-GCN with dual-stream architecture) and an adaptive grasping module integrating VLM-inferred target forces with MPC-based online optimization. The system is evaluated on three robot platforms (RMC-DA, YuMi, Unitree H1) using Sign and CSL-Daily datasets for retargeting accuracy, and on eight everyday objects for grasping performance. The retargeting module is compared against three baselines (NLO, VMR, ATP), and the adaptive grasping module is shown to improve grasp success rates from an average of 5.13/10 to 9.13/10 across tested objects.

Significance. The paper addresses a practically important problem: achieving cross-platform motion retargeting without paired human-robot data, combined with adaptive force control for compliant grasping. The dual-stream SAG-GCN architecture and the integration of VLM-based force estimation with MPC represent a reasonable system-level contribution. The authors provide quantitative retargeting metrics across multiple platforms (Tables I-IV), ablation studies (Table III), and grasping success rate comparisons (Table VI). The system is validated in both simulation and real-world environments, and the project page reference suggests reproducibility resources may be available.

major comments (1)
  1. §III-C, Eqs. (10)-(11): The adaptive grasping module's core optimization is formulated as gradient descent (Adam) on L(θ₁) = ||M(θ₁, θ₂) − F_target||² + λ||θ₁ − θ_prior₁||², where M is explicitly stated to be a random forest regressor. Random forests produce piecewise-constant outputs; their gradient with respect to continuous inputs is zero almost everywhere and undefined at split boundaries. The paper states that M serves as 'a differentiable surrogate for the subsequent gradient-based optimization,' but provides no mechanism (e.g., soft trees, finite-difference approximation, smoothed ensemble) that would make this true. If ∂L/∂θ₁ is effectively zero almost everywhere, the Adam updates in Eq. (11) cannot meaningfully steer joint commands toward the target force, and the optimization reduces to the regularization term λ||θ₁ − θ_prior₁||². This is load-bearing for the central claim of '
minor comments (8)
  1. §III-C: The VLM target forces (e.g., 300g for a water bottle, 30g for a paper cup) are stated as examples but the paper does not describe how these values were determined or validated. A brief explanation of the VLM prompting strategy and whether force values generalize across object instances with different masses would strengthen the presentation.
  2. §IV-A: The Sign and CSL-Daily datasets are sign language datasets. While they do contain diverse upper-body motions, the paper should briefly justify why sign language motions are appropriate proxies for manipulation-relevant teleoperation tasks, or acknowledge this as a limitation.
  3. Table II: The ATP MPJPE value of 0.9381 appears anomalously large compared to other entries (~0.09-0.10). Please verify this value is correct.
  4. §IV-E: The grasping evaluation uses 10 trials per object. While sufficient for a preliminary demonstration, a note acknowledging the limited sample size would be appropriate.
  5. §III-B, Eq. (3): The loss weights (λ_ee=1000, λ_ori=100, etc.) are stated without justification. A brief note on how these were selected would help reproducibility.
  6. §IV-E: The VLM used is 'Doubao-seed-1-6-vision-250815.' Please confirm this model name and provide a citation or access reference.
  7. Fig. 8: The force curves are described for five fingers, but the y-axis units and target force line are not clearly labeled. Adding these would improve clarity.
  8. §IV-D: The real-time performance of ~10 FPS is reported. It would be useful to note whether this includes the adaptive grasping module's latency or only the retargeting pipeline.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the careful reading and the constructive assessment. The referee raises one major technical concern regarding the differentiability of the random forest regressor used in the adaptive grasping module's MPC optimization loop. We agree that this is a valid and important point that requires correction in the manuscript. Below we address this comment in detail and describe the revisions we will make.

read point-by-point responses

  1. Referee: §III-C, Eqs. (10)-(11): The adaptive grasping module's core optimization is formulated as gradient descent (Adam) on L(θ₁) = ||M(θ₁, θ₂) − F_target||² + λ||θ₁ − θ_prior₁||², where M is explicitly stated to be a random forest regressor. Random forests produce piecewise-constant outputs; their gradient with respect to continuous inputs is zero almost everywhere and undefined at split boundaries. The paper states that M serves as 'a differentiable surrogate for the subsequent gradient-based optimization,' but provides no mechanism (e.g., soft trees, finite-difference approximation, smoothed ensemble) that would make this true. If ∂L/∂θ₁ is effectively zero almost everywhere, the Adam updates in Eq. (11) cannot meaningfully steer joint commands toward the target force, and the optimization reduces to the regularization term λ||θ₁ − θ_prior₁||². This is load-bearing for the central claim of '

    Authors: The referee is correct on the mathematical point. A standard random forest regressor produces piecewise-constant outputs, and its gradient with respect to continuous inputs is zero almost everywhere and undefined at split boundaries. The manuscript's description of M as 'a differentiable surrogate for the subsequent gradient-based optimization' is inaccurate as written, and the referee is also correct that if ∂L/∂θ₁ were effectively zero almost everywhere, the Adam updates in Eq. (11) would reduce to the regularization term alone, which would not explain the empirical results shown in Tables V and VI and Figure 8. We must therefore clarify what actually happens in the implementation and correct the manuscript accordingly. In our actual implementation, the online optimization does not rely on analytical gradients of the random forest. Instead, we use a finite-difference approximation: at each control cycle, the predicted force M(θ₁, θ₂) is evaluated at the current command θ₁ and at small perturbations θ₁ + δeᵢ along each dimension, and the resulting empirical gradient is fed into the Adam update. This is a standard black-box optimization approach that works with non-differentiable surrogates. The manuscript failed to describe this mechanism, which is a significant omission. We will revise §III-C to: (1) remove the incorrect claim that the random forest is 'a differentiable surrogate'; (2) explicitly describe the finite-difference gradient estimation procedure used in the Adam optimization loop; (3) add the perturbation step size δ and related implementation details; and (4) clarify that the random forest serves as a fast non-analytic forward predictor whose gradients are approximated numerically rather than computed in closed form. We acknowledge that the current wording revision: yes

Circularity Check

0 steps flagged

No significant circularity; self-citation to prior VMR work is comparative, not load-bearing for the central claims.

full rationale

The paper's two main contributions — the SAG-GCN motion retargeting module and the VLM+MPC adaptive grasping module — are each presented with independent architectural descriptions, loss functions, and experimental evaluations against external baselines (NLO [35], ATP [38]) and ablation studies. The self-citation to VMR [37] (Lai, Ju, Gao — overlapping authors) appears only as one of several comparison baselines in Tables I–II, not as a foundational premise that the current method's correctness depends on. The SAG-GCN architecture is described in full (Eqs. 4–7, Figs. 3–4) without requiring the reader to accept results from [37]. The adaptive grasping module's formulation (Eqs. 8–11) is self-contained. The reader's concern about the random forest's differentiability is a correctness/implementation issue, not a circularity issue — the paper does not define the surrogate model in terms of the target result it claims to predict. The VLM force values (e.g., 300g for a water bottle) are stated as examples of VLM output, not fitted parameters renamed as predictions. No step in the derivation chain reduces to its own inputs by construction. The modest improvement margins over VMR and the architectural overlap with prior work are legitimate concerns for novelty and correctness risk, but they do not constitute circularity as defined here. Score 2 reflects the presence of self-citation that is comparative rather than load-bearing, which is minor and does not undermine the independence of the central claims.

Axiom & Free-Parameter Ledger

9 free parameters · 4 axioms · 0 invented entities

The paper introduces no new physical entities, particles, or forces. The SAG-GCN, SBB, and GRB are architectural modules, not postulated physical entities. The free parameters are loss weights and learning rates that are set by hand. The domain assumptions concern dataset choice, VLM force inference validity, random forest differentiability, and sim-to-real transfer.

free parameters (9)
  • λ_ee = 1000
    End-effector position loss weight in Eq. 3, set by hand without justification
  • λ_ori = 100
    End-effector orientation loss weight in Eq. 3, set by hand
  • λ_norm = 1000
    Arm normal vector loss weight in Eq. 3, set by hand
  • λ_d = 1000
    Dynamics loss weight in Eq. 3, set by hand
  • λ_fin1 = 100
    Fingertip orientation loss weight in Eq. 3, set by hand
  • λ_fin2 = 100
    Finger angle loss weight in Eq. 3, set by hand
  • λ (MPC regularization) = not specified
    Regularization weight in Eq. 10 balancing force tracking and joint command smoothness, value not given
  • Learning rate = 1e-4
    Adam optimizer learning rate for SAG-GCN training
  • VLM target forces = 30-300g range
    Object-specific target grasping forces (e.g., 300g for water bottle, 30g for paper cup) inferred by VLM, treated as ground truth without validation
axioms (4)
  • domain assumption Sign language motion datasets (Sign, CSL-Daily) are adequate proxies for evaluating teleoperation retargeting quality
    §IV-A: The retargeting evaluation uses sign language datasets, but teleoperation involves manipulation motions. The paper does not justify why sign language motions are representative of teleoperation tasks.
  • domain assumption VLM-inferred grasping forces are appropriate target references for compliant grasping
    §III-C: The VLM outputs target forces like '300g for a water bottle' which are used as reference values for MPC optimization. No validation of these force values is provided.
  • domain assumption A random forest regressor can accurately model the joint angle-to-force mapping for a dexterous hand
    §III-C: The force prediction model M is a random forest trained on historical data, used as a differentiable surrogate. Random forests are not natively differentiable, and the paper does not explain how gradients are computed through this model.
  • domain assumption Simulation results on YuMi and Unitree H1 transfer to real-world deployment
    §IV-A: Two of three robot platforms are tested only in simulation. The paper claims cross-platform generalization but does not validate on physical YuMi or Unitree H1 robots.

pith-pipeline@v1.1.0-glm · 15216 in / 3257 out tokens · 618871 ms · 2026-07-08T21:31:15.378816+00:00 · methodology

0 comments
read the original abstract

In dual-arm dexterous teleoperation, cross-platform generalization of motion retargeting and interactivity of grasping are crucial. However, the heterogeneity of robotic architectures and the wide variety of grasping objects pose significant challenges to achieving precise motion retargeting and compliant grasping in dual-arm dexterous teleoperation. To address these challenges, a dual-arm dexterous teleoperation system (DexTele) is proposed based on motion retargeting and adaptive force control. First, a vision-based motion retargeting module is designed to generate preliminary robot motions from human images. In this module, a motion-graph encoder and latent optimization are proposed for precise and convenient cross-platform motion retargeting. Second, an adaptive grasping module is designed to achieve compliant grasping. This module combines a vision-language model (VLM) with model predictive control (MPC), allowing the system to predict the required grasping force for a target object and perform gradient-based online optimization. Finally, extensive experiments demonstrate that the DexTele achieves precise motion retargeting and compliant grasping with generalization across multiple robot platforms.

Figures

Figures reproduced from arXiv: 2607.05883 by Junjie Hu, Qing Gao, Xianfeng Cheng, Yuanchuan Lai, Zhaojie Ju, Ziyan Liang.

Figure 1
Figure 1. Figure 1: Schematic of the teleoperation system. Part (a) illustrates previous [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the dual-arm dexterous teleoperation system. Human motions are first captured via FrankMocap and processed into 3D body and [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 5
Figure 5. Figure 5: Pipeline of the Adaptive Grasping Module. When the dexterous [PITH_FULL_IMAGE:figures/full_fig_p004_5.png] view at source ↗
Figure 4
Figure 4. Figure 4: Structures of the Spatial Basic Block and Gated Residual Block. [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Demonstration of motion retargeting across multiple robot plat [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Illustration of the grasping results on eight different object [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Grasping force curves for the water bottle. The five colored lines [PITH_FULL_IMAGE:figures/full_fig_p007_8.png] view at source ↗

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

38 extracted references · 38 canonical work pages · 3 internal anchors

  1. [1]

    S. Li, N. Hendrich, H. Liang, et al., ”A dexterous hand-arm teleoper- ation system based on hand pose estimation and active vision,” IEEE Trans. Cybern., vol. 54, no. 3, pp. 1417–1428, 2022

    work page 2022

  2. [2]

    Robotic Telekinesis: Learning a Robotic Hand Imitator by Watching Humans on Youtube

    A. Sivakumar, K. Shaw, and D. Pathak, ”Robotic telekinesis: Learning a robotic hand imitator by watching humans on YouTube,” arXiv preprint, arXiv:2202.10448, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  3. [3]

    Z. Fu, T. Z. Zhao, and C. Finn, ”Mobile Aloha: Learning bimanual mobile manipulation with low-cost whole-body teleoperation,” arXiv preprint arXiv:2401.02117, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  4. [4]

    Open-TeleVision: Teleoperation with Immersive Active Visual Feedback

    X. Cheng, J. Li, S. Yang, et al., ”Open-television: Teleoperation with immersive active visual feedback,” arXiv preprint arXiv:2407.01512, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  5. [5]

    C. Zeng, S. Li, Y . Jiang, and others, ”Learning compliant grasping and manipulation by teleoperation with adaptive force control,” in Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst., 2021, pp. 717-724

    work page 2021

  6. [6]

    Li, et al., ”Vision-based teleoperation of shadow dexterous hand using end-to-end deep neural network,” in Proc

    S. Li, et al., ”Vision-based teleoperation of shadow dexterous hand using end-to-end deep neural network,” in Proc. Int. Conf. Robot. Automat., 2019, pp. 416–422

    work page 2019

  7. [7]

    S. Baek, A. Kim, J. Y . Choi, et al., ”Human motion retargeting to a full-scale humanoid robot using a monocular camera and human pose estimation,” Int. J. Control, Autom. Syst., vol. 22, no. 9, pp. 2860–2870, 2024

    work page 2024

  8. [8]

    C. Zeng, S. Li, Z. Chen, et al., ”Multifingered robot hand compliant manipulation based on vision-based demonstration and adaptive force control,” IEEE Trans. Neural Netw. Learn. Syst., vol. 34, no. 9, pp. 5452–5463, 2022

    work page 2022

  9. [9]

    TeachMe: Three-phase learning framework for robotic motion imitation based on interactive teaching and rein- forcement learning,

    T. Kim and J.-H. Lee, “TeachMe: Three-phase learning framework for robotic motion imitation based on interactive teaching and rein- forcement learning,” in Proc. IEEE Int. Conf. Robot Hum. Interact. Commun., Oct. 2019, pp. 1–8

    work page 2019

  10. [10]

    Patel, T

    S. Patel, T. Garg, G. Patel, et al., ”Motion retargeting and machine learning for humanoid robotics,” in Proc. 2020 International Sympo- sium on Devices, Circuits and Systems (ISDCS), IEEE, 2020, pp. 1–5

    work page 2020

  11. [11]

    Z. Deng, Y . Jonetzko, L. Zhang, et al., ”Grasping force control of multi-fingered robotic hands through tactile sensing for object stabilization,” Sensors, vol. 20, no. 4, p. 1050, 2020

    work page 2020

  12. [12]

    Q. Tang, H. Yang, W. Wang, et al., ”Grasp compliant control using adaptive admittance control methods for flexible objects,” in Proc. International Conference on Intelligent Robotics and Applications, Singapore, 2023, pp. 515–525

    work page 2023

  13. [13]

    Cortinovis, G

    S. Cortinovis, G. Vitrani, M. Maggiali, et al., ”Control methodologies for robotic grippers: A review,” Actuators, vol. 12, no. 8, p. 332, 2023

    work page 2023

  14. [14]

    B. Fang, F. Sun, H. Liu, and C. Liu, ”3D human gesture capturing and recognition by the IMMU-based data glove,” Neurocomputing, vol. 277, pp. 198–207, Feb. 2018

    work page 2018

  15. [15]

    D. Shi, S. Jin, C. Yang, et al., ”Exploring the Synergistic Effects of Teleoperation Scaling Ratio and Learning from Demonstration,” IEEE Trans. Autom. Sci. Eng., 2025

    work page 2025

  16. [16]

    Zhang, Q

    M. Zhang, Q. Gao, Y . Lai, et al., ”HR-GCN: 2D-3D Whole-body Pose Estimation with High-Resolution Graph Convolutional Network From a Monocular Camera,” IEEE Sens. J., 2025

    work page 2025

  17. [17]

    Zeng, et al., ”Learning compliant grasping and manipulation by teleoperation with adaptive force control,” in Proc

    C. Zeng, et al., ”Learning compliant grasping and manipulation by teleoperation with adaptive force control,” in Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst., Sep. 2021, pp. 717–724

    work page 2021

  18. [18]

    L. S. Li, J. Jiang, P. Ruppel, et al., ”A mobile robot hand-arm teleoperation system by vision and IMU,” in Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst., 2020, pp. 10900–10906

    work page 2020

  19. [19]

    Lu, et al., ”Mobile-TeleVision: Predictive Motion Priors for Hu- manoid Whole-Body Control,” in Proc

    C. Lu, et al., ”Mobile-TeleVision: Predictive Motion Priors for Hu- manoid Whole-Body Control,” in Proc. 2025 IEEE Int. Conf. Robotics and Automation, Atlanta, GA, USA, 2025, pp. 5364–5371

    work page 2025

  20. [20]

    Z. Yang, S. Bien, S. Nertinger, et al., ”An optimization-based scheme for real-time transfer of human arm motion to robot arm,” in Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst., 2024, pp. 12220–12225

    work page 2024

  21. [21]

    C-3PO: Cyclic-three-phase optimization for human-robot motion retargeting based on reinforcement learning,

    T. Kim and J.-H. Lee, “C-3PO: Cyclic-three-phase optimization for human-robot motion retargeting based on reinforcement learning,” inProc.IEEE Int. Conf. Robot. Automat., May 2020, pp. 8425–8432

    work page 2020

  22. [22]

    S. Choi, M. J. Song, H. Ahn, and J. Kim, ”Self-supervised motion retargeting with safety guarantee,” in Proc. IEEE Int. Conf. Robot. Automat., May 2021, pp. 8097–8103

    work page 2021

  23. [23]

    Liang, et al., ”PointNetGPD: Detecting grasp configurations from point sets,” in Proc

    H. Liang, et al., ”PointNetGPD: Detecting grasp configurations from point sets,” in Proc. IEEE Int. Conf. Robot. Automat., May 2019, pp. 3629–3635

    work page 2019

  24. [24]

    Q. Lu, M. Van der Merwe, B. Sundaralingam, and T. Hermans, ”Multifingered grasp planning via inference in deep neural networks: Outperforming sampling by learning differentiable models,” IEEE Robot. Autom. Mag., vol. 27, no. 2, pp. 55–65, Jun. 2020

    work page 2020

  25. [25]

    Wimbock, C

    T. Wimbock, C. Ott, and G. Hirzinger, ”Analysis and experimental evaluation of the intrinsically passive controller (IPC) for multifingered hands,” in Proc. IEEE Int. Conf. Robot. Automat., May 2008, pp. 278–284

    work page 2008

  26. [26]

    M. Li, H. Yin, K. Tahara, and A. Billard, ”Learning object-level impedance control for robust grasping and dexterous manipulation,” in Proc. IEEE Int. Conf. Robot. Automat., May 2014, pp. 6784–6791

    work page 2014

  27. [27]

    Xu and Y

    Z. Xu and Y . She, ”LeTac-MPC: Learning model predictive control for tactile-reactive grasping,” IEEE Trans. Robot., 2024

    work page 2024

  28. [28]

    L. Shi, C. Mucchiani, and K. Karydis, ”Online modeling and control of soft multi-fingered grippers via Koopman operator theory,” in Proc. IEEE Int. Conf. Autom. Sci. Eng., 2022, pp. 1946–1952

    work page 2022

  29. [29]

    D. Tian, X. Lin, and Y . Sun, ”Adaptive motion planning for multi- fingered functional grasp via force feedback,” in Proc. IEEE-RAS Int. Conf. Humanoid Robots, 2024, pp. 835–842

    work page 2024

  30. [30]

    Y . Rong, T. Shiratori, and H. Joo, ”Frankmocap: A monocular 3D whole-body pose estimation system via regression and integration,” in Proc. IEEE/CVF Int. Conf. Comput. Vis., 2021, pp. 1749–1759

    work page 2021

  31. [31]

    H. Zhou, W. Zhou, W. Qi, et al., ”Improving sign language translation with monolingual data by sign back-translation,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2021, pp. 1316–1325

    work page 2021

  32. [32]

    Fey and J

    M. Fey and J. E. Lenssen, ”Fast graph representation learning with PyTorch geometric,” in Proc. Int. Conf. Learn. Representations Work- shop Representation Learn. Graphs Manifolds, May 2019

    work page 2019

  33. [33]

    Villegas, J

    R. Villegas, J. Yang, D. Ceylan, et al., ”Neural kinematic networks for unsupervised motion retargetting,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit., 2018, pp. 8639–8648

    work page 2018

  34. [34]

    Greer, N

    R. Greer, N. Deo, and M. Trivedi, ”Trajectory prediction in au- tonomous driving with a lane heading auxiliary loss,” IEEE Robot. Autom. Lett., vol. 6, no. 3, pp. 4907–4914, 2021

    work page 2021

  35. [35]

    Zhang, et al., ”Kinematic motion retargeting via neural latent optimization for learning sign language,” IEEE Robot

    H. Zhang, et al., ”Kinematic motion retargeting via neural latent optimization for learning sign language,” IEEE Robot. Autom. Lett., vol. 7, no. 2, pp. 4582–4589, Apr. 2022

    work page 2022

  36. [36]

    J. Li, S. Bian, C. Xu, et al., ”D&D: Learning human dynamics from dynamic camera,” in Proc. Eur. Conf. Comput. Vis., Cham: Springer, 2022, pp. 479–496

    work page 2022

  37. [37]

    Motion retargeting using graph neural network for vision-guided dexterous robot teleoperation,

    Y . Lai, Z. Ju, and Q. Gao, “Motion retargeting using graph neural network for vision-guided dexterous robot teleoperation,” in Proc. 17th Int. Conv. Rehabil. Eng. Assistive Technol. (i-CREATe), 2024, pp. 1–6

    work page 2024

  38. [38]

    Qin, et al., ”AnyTeleop: A General Vision-Based Dexterous Robot Arm-Hand Teleoperation System,” in Proc

    Y . Qin, et al., ”AnyTeleop: A General Vision-Based Dexterous Robot Arm-Hand Teleoperation System,” in Proc. Robotics: Science and Systems (RSS), 2023

    work page 2023