Pith. sign in

REVIEW 4 major objections 5 minor 42 references

DexDirect claims that letting operators physically guide a robot arm while a webcam tracks the fingers makes dexterous demonstration collection 3.2x to 17.2x more efficient than low-setup alternatives.

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

2026-08-01 01:27 UTC pith:453UOVEG

load-bearing objection A genuinely useful interface idea with a real user study, but the headline 17.2x throughput gain is probably inflated by a shaky AnyTeleop baseline and short practice; worth reviewing, not worth taking at face value. the 4 major comments →

arxiv 2607.27784 v1 pith:453UOVEG submitted 2026-07-30 cs.RO

DexDirect: Direct Kinesthetic Arm Guidance for Efficient Dexterous Demonstration Collection

Beom Jun Kim , Shiu-Jen Wang , Jonathan Liu , Alvin Zhu , Quanyou Wang , Hanzhang Fang , Feng Xu , Mingzhang Zhu
show 2 more authors
Yuchen Cui Dennis W. Hong
This is my paper
classification cs.RO
keywords dexterous manipulationkinesthetic teachingdemonstration collectionteleoperation interfacevision-based hand retargetingdiffusion policyuser studyrobot learning
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 argues that the main bottleneck in dexterous robot learning is collecting enough good demonstrations, and that existing low-cost interfaces—purely vision-based or phone-based tracking—force operators to fight latency, occlusions, and coordinate remapping. DexDirect sidesteps that proxy loop for the arm: the operator drags the gravity-compensated robot arm by a handle, so arm motion is read directly from joint encoders, while a single webcam retargets the operator's free hand to the robot hand's fingers. In a ten-participant, five-task user study it collected 481 successful demonstrations versus 151 and 28 for the two baselines, with the highest success rate on every task and lower perceived workload except for physical demand. The paper concludes that direct arm guidance plus vision-based finger retargeting is a low-setup, scalable interface for dexterous demonstration collection, and shows that a diffusion policy trained on its data reaches 90% success on a cube pick-and-place task.

Core claim

Central discovery: separating arm control from finger control—direct physical command of the arm, webcam-only finger retargeting—removes the tracking, retargeting, and coordinate-mapping burden that dominates low-setup dexterous teleoperation. Evidence comes from a ten-participant, five-task user study: DexDirect collected 481 successful demonstrations, compared with 151 for the phone-based baseline and 28 for the vision-only baseline, with the best success rate on every task and reduced perceived workload except physical demand. A diffusion policy trained on 200 DexDirect demonstrations reached 90% success on cube pick-and-place, showing the data supports downstream imitation learning.

What carries the argument

The mechanism is a hybrid control split. The 6-DoF arm is commanded by direct kinesthetic guidance in a gravity-compensated, zero-stiffness drag mode; joint encoders record the arm trajectory, so no pose estimation, inverse kinematics, or retargeting is needed for the arm. The 16-joint, 13-DoF robot hand is commanded by a single palm-facing webcam: a hand-tracking model produces wrist-relative landmarks, and an optimization-based retargeting step maps those to robot finger vectors while enforcing joint limits. A rigid wrist handle transmits passive contact cues to the operator, and a common clock aligns arm, hand, tactile, and wrist-camera streams for policy learning.

Load-bearing premise

The paper's throughput ratios depend on the assumption that its in-house AnyTeleop and TeleDex reproductions (Section IV-A) faithfully represent the published systems; the 0% success rates for AnyTeleop on two tasks and the absence of validation against published results mean the 17.2x headline could be inflated if those baselines were under-powered.

What would settle it

Run the same five-task, fixed-time-budget protocol using the original AnyTeleop and TeleDex implementations (or their published evaluation setups) instead of in-house reproductions, and compare per-task success counts; if the gap falls below roughly 2x for TeleDex or AnyTeleop achieves non-zero success on whiteboard erasing and keyboard typing, the central claim of DexDirect's efficiency advantage is not supported.

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

If this is right

  • A low-cost dexterous data-collection interface can be built from a gravity-compensated arm and a commodity webcam, without leader arms, exoskeletons, or operator-specific calibration.
  • Because arm trajectories are physically attained robot configurations, the collected demonstrations are directly executable and do not require offline re-execution or visual inpainting for policy training.
  • The workload profile—mental/cognitive load traded for physical effort—points to an ergonomic design target for future interfaces: reducing physical demand while keeping the direct-coupling advantage.
  • The 90% success of a diffusion policy trained purely on DexDirect demonstrations suggests the interface's data quality is sufficient for modern imitation learning, at least for short-horizon pick-and-place.
  • The per-task completion-time speedups (1.8x to 5.6x) show faster demonstration collection even when the operator is already proficient, not just a higher success rate.

Where Pith is reading between the lines

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

  • The hybrid split implies an explicit trade-off curve for teleoperation interfaces: from direct physical coupling (low cognitive load, high physical load, no remote operation) to pure tracking (high cognitive load, low physical load, remote-capable); the ideal operating point likely depends on arm inertia and task contact richness.
  • Because the vision pathway only handles wrist-relative finger geometry, the same hand-retargeting module could in principle be reused with any kinesthetic-guidance arm, making DexDirect a modular recipe rather than a single robot setup.
  • A natural stress test is to scale to a larger, higher-inertia arm: the physical-demand penalty could overtake the cognitive savings, so the throughput advantage may not transfer to full-size manipulators.
  • The paper does not isolate contributions of direct coupling versus reduced latency versus passive contact cues; an ablation that adds artificial delay or removes the handle would quantify which part of the 3.2x/17.2x gain is actually due to kinesthetic guidance.

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

4 major / 5 minor

Summary. DexDirect is a hybrid demonstration-collection interface in which the operator physically guides a gravity-compensated 6-DoF robot arm through a wrist-mounted handle while a webcam-based retargeting pipeline maps the operator's other hand to a 16-joint/13-DoF dexterous hand. The arm state is read directly from joint encoders; finger articulation uses MediaPipe and the dex-retargeting framework released with AnyTeleop. In a controlled user study with ten participants, five contact-rich tasks, and three interfaces, the paper reports 481 successful demonstrations for DexDirect versus 151 for TeleDex and 28 for AnyTeleop, corresponding to 3.2x and 17.2x throughput advantages. An adapted NASA-TLX shows lower overall workload for DexDirect, and a diffusion policy trained on 200 DexDirect demonstrations achieves 90% success on a cube pick-and-place task. The authors conclude that direct kinesthetic arm guidance combined with vision-based hand retargeting provides an efficient, low-setup, scalable interface for dexterous demonstration collection.

Significance. The interface idea is simple, well-motivated, and plausible: replacing the arm-side proxy loop with direct physical guidance addresses a real bottleneck in dexterous demonstration collection while keeping the finger side commodity-vision-based. The study has genuine strengths: a within-participant design with counterbalanced ordering, fixed robot/hand/finger-retargeting pipeline across conditions, explicit task success criteria, and a downstream policy-learning check. If the throughput advantage survives more rigorous baseline validation, DexDirect would be a practically useful contribution. However, the headline quantitative claims rest on two assumptions that are not fully secured: that the reproduced AnyTeleop baseline is representative of the published system, and that the one-minute practice protocol is fair to all interfaces. The manuscript is candid in its limitations section, but that section does not identify the practice-time confound that most directly affects the central comparison.

major comments (4)
  1. [IV-A, Table I] The AnyTeleop baseline is an in-house reproduction that is not validated against the published system. AnyTeleop scores 0.00 on whiteboard erasing and keyboard typing; the post-hoc video review attributes this to depth regulation and latency, but that does not distinguish an inherent method limitation from an artifact of the reproduction (RealSense placement, hand plate, CuRobo tuning, controller gains, or operator practice). Without evidence that the reproduced pipeline performs comparably to the published implementation on its own tasks or a shared benchmark, the 17.2x ratio is not robust. Please provide validation against publicly available AnyTeleop results, detailed calibration/setup information, or a side-by-side comparison.
  2. [IV-B, IV-C] The headline throughput numbers (481 vs. 151 vs. 28 successful demonstrations) are aggregate point estimates with no per-participant distribution, confidence intervals, or paired significance tests. With N=10, one or two participants could drive much of the ratio. Report per-participant counts, medians/IQR or bootstrap confidence intervals, and a paired test (e.g., Wilcoxon on counts or log throughput). The same applies to the per-task success-rate columns in Table I, where only SEM is reported.
  3. [IV-B, V] Participants receive only one minute of practice per task per interface before timed trials. Direct kinesthetic guidance is likely intuitive immediately, whereas AnyTeleop and TeleDex require learning coordinate mappings and compensating for latency; one minute therefore plausibly understates baseline steady-state throughput and inflates the 3.2x/17.2x ratios. Section V lists limitations but does not identify this confound. Please justify the practice duration empirically—for example, show that success rates plateau within the practice window, compare first versus second halves of the timed session, or include a longer-practice control condition.
  4. [IV-D, Fig. 5] The abstract and contribution list claim that DexDirect 'greatly reduces mental demand, effort, and frustration,' but statistical tests are reported only for the overall Raw TLX score. The dimension-level comparisons are presented as mean differences with 95% confidence intervals. With ten participants and six dimensions, uncorrected tests on individual dimensions may not support the strength of the claim; please report planned dimension-level analyses or explicitly reframe these as descriptive observations.
minor comments (5)
  1. [IV-B] Typo: 'basline methods' should be 'baseline methods.'
  2. [III-C] The sentence describing use of the dex-retargeting framework is a bit buried; state explicitly that the finger-reargeting module is the same as AnyTeleop's so readers can isolate the arm-interface contribution.
  3. [IV-E] The policy evaluation is a useful feasibility check, but 20 rollouts on one task with one operator's demonstrations is too small to support general claims. Please label it as a feasibility result and report initial-state randomization details and per-seed variability if available.
  4. [Table I] The dash for AnyTeleop completion time is understandable, but consider adding 'no successful trials' in the caption rather than only in the footnote; also clarify whether SEM is over participants or attempts.
  5. [V] The limitations section already acknowledges combined-effect confounds and single-platform scope; consider adding the practice-time issue there as well, since it is directly measurable and relevant to the central claim.

Circularity Check

0 steps flagged

No circularity: the comparison is empirical and the claimed throughput/policy results are not fitted to or derived from the interface design.

full rationale

The paper makes no formal derivation whose outputs are equivalent to its inputs. The central claims—higher demonstration throughput, lower workload, and 90% downstream policy success—are obtained from a held-out user study (10 participants, 5 tasks, fixed time budgets) and from an independently collected 200-demonstration policy-training set. The finger-retargeting pipeline is held fixed across conditions, so the interface comparison is not a fitted-input-called-prediction loop; the policy success rate is a downstream check, not a parameter used to fit the interface. Self-citations appear only as hardware/system references (MIDAS hand [36], DexExo [9]) or as a baseline whose open-source software was used (TeleDex [12]), and none of these is invoked as a uniqueness theorem or ansatz that forces the paper's conclusions. The limitation statement that the evaluation does not isolate coordinate remapping, latency, and passive contact cues is an acknowledged confound, not a circular reduction. Concerns about baseline representativeness or the one-minute practice window are correctness/validity risks, not circularity, and are not the kind of same-by-construction equivalence this pass targets. No specific equation, definition, or fit can be exhibited that reduces a claimed result to its own input.

Axiom & Free-Parameter Ledger

1 free parameters · 4 axioms · 0 invented entities

The paper makes no derived predictions or fitted mathematical models; its load-bearing assumptions are about hardware behavior, sensor fidelity, and baseline fairness. The only explicit fitted numbers are gravity-compensation calibration constants, which enable operation but are not used to predict the outcome. The strongest non-empirical risk is the assumption that baselines perform as well as their original publications, which the paper does not validate.

free parameters (1)
  • Gravity-compensation system-identification parameters = not reported (gravity direction, payload mass/COM, per-joint torque scaling, Coulomb friction)
    Fitted per setup in Section III-B to enable the zero-stiffness drag mode. They are operationally necessary but do not determine the comparison outcome; values are not published, limiting replication.
axioms (4)
  • domain assumption Kinesthetic arm guidance produces accurate, executable arm trajectories and useful passive contact cues for the operator.
    Central to the interface design; supported by prior kinesthetic teaching literature ([14-16]) but not independently verified here; discussed in Sec. III-B and Sec. V.
  • domain assumption Webcam-based MediaPipe landmarks plus AnyTeleop's dex-retargeting optimizer provides sufficient finger-control fidelity for contact-rich tasks.
    The finger pathway is lifted from prior work ([37],[10]) with no reported tracking or retargeting error quantification; the authors list this as a limitation in Sec. V.
  • ad hoc to paper The reproduced AnyTeleop and TeleDex baselines represent each method's typical performance.
    The entire throughput comparison rests on this; AnyTeleop's 0% success on two tasks and lack of validation against published baseline results make it a potential source of bias (Sec. IV-A, Table I).
  • domain assumption Ten participants, five tasks, and one robot platform generalize to broader dexterous demonstration collection.
    Sample scope is stated in Sec. V as a limitation; the 'scalable' conclusion in the abstract goes beyond the evidence.

pith-pipeline@v1.3.0-daily-deepseek · 9518 in / 14765 out tokens · 144294 ms · 2026-08-01T01:27:14.638507+00:00 · methodology

0 comments
read the original abstract

Scalable collection of dexterous manipulation demonstrations remains a major bottleneck for robot learning. High-fidelity interfaces often require costly hardware and extensive setup, while low-setup, low cost alternatives tend to provide less precise control and impose greater cognitive workload on operators. We present DexDirect, a direct kinesthetic arm guidance for efficient dexterous demonstration collection. The operator drags a 6-DoF gravity-compensated robot arm directly by a handle, while a single webcam retargets operator's other hand onto a 16 joints 13-DoF dexterous robot hand. User studies suggest DexDirect collects 17.2x and 3.2x more successful demonstrations compared to purely vision (AnyTeleop) and pose-tracking (TeleDex) baselines. An adapted NASA-TLX shows DexDirect greatly reduces mental demand, effort, and frustration, despite raising physical demand. A diffusion policy trained on DexDirect demonstrations reaches a 90% success rate on a cube pick-and-place task. These results suggest that direct kinesthetic arm guidance combined with vision-based hand retargeting provides an efficient low-setup and scalable interface for collecting dexterous manipulation demonstrations

Figures

Figures reproduced from arXiv: 2607.27784 by Alvin Zhu, Beom Jun Kim, Dennis W. Hong, Feng Xu, Hanzhang Fang, Jonathan Liu, Mingzhang Zhu, Quanyou Wang, Shiu-Jen Wang, Yuchen Cui.

Figure 1
Figure 1. Figure 1: DexDirect combines direct kinesthetic guidance of the robot [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: System overview of DexDirect The operator directly guides the gravity-compensated robot arm while controlling the dexterous hand through vision-based hand-pose retargeting. The collected joint, visual, and tactile observations are synchronized and used to train a multimodal diffusion policy. At inference, the policy predicts robot action chunks conditioned on the current camera and tactile observations [P… view at source ↗
Figure 3
Figure 3. Figure 3: Evaluation tasks. DexDirect is evaluated on five contact-rich dexterous manipulation tasks: Jenga manipulation, USB unplugging, Lazy Susan pick-and-place, whiteboard erasing, and keyboard typing. DexDirect TeleDex AnyTeleop [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of operator interfaces for dexterous teleop￾eration. DexDirect combines direct physical guidance of the robot arm with vision-based hand retargeting, whereas TeleDex relies on a handheld control interface and AnyTeleop tracks the operator’s arm and hand motion entirely through vision. effector poses are obtained through forward kinematics. No arm-side pose estimation, retargeting, inverse-kinema… view at source ↗
Figure 5
Figure 5. Figure 5: Mean adapted NASA–TLX scores for AnyTeleop, TeleDex, [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Representative policy rollouts. The learned policy grasps a cube from a randomized initial pose, lifts it, and places it onto the target box. Each sequence shows the rollout at t = {0, 5, 10, 15} s. this trade is favorable for repeated demonstration collection. E. Policy Training To evaluate whether DexDirect is suitable for end-to-end policy learning, we design a simple cube pick-and-place task. The robot… 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

42 extracted references · 21 linked inside Pith

  1. [1]

    Universal manipulation interface: In-the-wild robot teaching without in-the-wild robots,

    C. Chi et al., “Universal manipulation interface: In-the-wild robot teaching without in-the-wild robots,” inRobotics: Science and Systems (RSS), 2024. arXiv:2402.10329 [cs.RO]

    Pith/arXiv arXiv 2024

  2. [2]

    Gello: A general, low-cost, and intuitive teleoperation framework for robot manipula- tors,

    P. Wu, Y . Shentu, Z. Yi, X. Lin, and P. Abbeel, “Gello: A general, low-cost, and intuitive teleoperation framework for robot manipula- tors,” inIEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2024. arXiv:2309.13037 [cs.RO]

    Pith/arXiv arXiv 2024

  3. [3]

    Learning fine-grained bimanual manipulation with low-cost hardware,

    T. Z. Zhao, V . Kumar, S. Levine, and C. Finn, “Learning fine-grained bimanual manipulation with low-cost hardware,” inRobotics: Sci- ence and Systems (RSS), 2023. arXiv:2304.13705 [cs.RO]

    Pith/arXiv arXiv 2023

  4. [4]

    Si et al.,Exostart: Efficient learning for dexterous manipulation with sensorized exoskeleton demonstrations, 2025

    Z. Si et al.,Exostart: Efficient learning for dexterous manipulation with sensorized exoskeleton demonstrations, 2025. arXiv:2506 . 11775 [cs.RO]

    2025

  5. [5]

    Dexpilot: Vision-based teleoperation of dexterous robotic hand-arm system,

    A. Handa et al., “Dexpilot: Vision-based teleoperation of dexterous robotic hand-arm system,” inIEEE International Conference on Robotics and Automation (ICRA), 2020

    2020

  6. [6]

    Dexumi: Using human hand as the universal manipu- lation interface for dexterous manipulation,

    M. Xu et al., “Dexumi: Using human hand as the universal manipu- lation interface for dexterous manipulation,” inConference on Robot Learning, PMLR, 2025, pp. 437–459

    2025

  7. [7]

    Ace: A cross-platform visual-exoskeletons system for low-cost dexterous teleoperation,

    S. Yang et al., “Ace: A cross-platform visual-exoskeletons system for low-cost dexterous teleoperation,” inConference on Robot Learning (CoRL), 2024. arXiv:2408.11805 [cs.RO]

    Pith/arXiv arXiv 2024

  8. [8]

    Fang et al.,Dexop: A device for robotic transfer of dexterous human manipulation, 2025

    H.-S. Fang et al.,Dexop: A device for robotic transfer of dexterous human manipulation, 2025. arXiv:2509.04441 [cs.RO]

    Pith/arXiv arXiv 2025

  9. [9]

    Zhu et al.,Dexexo: A wearability-first dexterous exoskeleton for operator-agnostic demonstration and learning, 2026

    A. Zhu et al.,Dexexo: A wearability-first dexterous exoskeleton for operator-agnostic demonstration and learning, 2026. arXiv:2603. 17323 [cs.RO]

    2026

  10. [10]

    Anyteleop: A general vision-based dexterous robot arm-hand teleoperation system,

    Y . Qin et al., “Anyteleop: A general vision-based dexterous robot arm-hand teleoperation system,” inRobotics: Science and Systems, 2023

    2023

  11. [11]

    Sivakumar, K

    A. Sivakumar, K. Shaw, and D. Pathak,Robotic telekinesis: Learn- ing a robotic hand imitator by watching humans on youtube, 2022. arXiv:2202.10448 [cs.RO]

    Pith/arXiv arXiv 2022

  12. [12]

    Rayyan, M

    O. Rayyan, M. Gilles, and Y . Cui,Teledex: Accessible dexterous teleoperation, 2026. arXiv:2603.17065 [cs.RO]

    arXiv 2026

  13. [13]

    Omnih2o: Universal and dexterous human-to-humanoid whole-body teleoperation and learning,

    T. He et al., “Omnih2o: Universal and dexterous human-to-humanoid whole-body teleoperation and learning,” inConference on Robot Learning (CoRL), 2024. arXiv:2406.08858 [cs.RO]

    Pith/arXiv arXiv 2024

  14. [14]

    A user study on kinesthetic teaching of redundant robots in task and configuration space,

    S. Wrede, C. Emmerich, R. Gr ¨unberg, A. Nordmann, A. Swadzba, and J. J. Steil, “A user study on kinesthetic teaching of redundant robots in task and configuration space,”Journal of Human-Robot Interaction, vol. 2, no. 1, pp. 56–81, 2013

    2013

  15. [15]

    Trajectories and keyframes for kinesthetic teaching: A human-robot interaction perspective,

    B. Akgun, M. Cakmak, J. W. Yoo, and A. L. Thomaz, “Trajectories and keyframes for kinesthetic teaching: A human-robot interaction perspective,” inProceedings of the ACM/IEEE International Confer- ence on Human-Robot Interaction (HRI), ACM, 2012, pp. 391–398

    2012

  16. [16]

    How to train your robots? the impact of demonstration modality on imitation learning,

    H. Li, Y . Cui, and D. Sadigh, “How to train your robots? the impact of demonstration modality on imitation learning,” inProceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2025, pp. 1113–1120

    2025

  17. [17]

    Development of nasa-tlx (task load index): Results of empirical and theoretical research,

    S. G. Hart and L. E. Staveland, “Development of nasa-tlx (task load index): Results of empirical and theoretical research,” inHuman Mental Workload, P. A. Hancock and N. Meshkati, Eds., North- Holland, 1988, pp. 139–183

    1988

  18. [18]

    Cadene et al.,Lerobot: State-of-the-art machine learning for real-world robotics in pytorch,https : / / github

    R. Cadene et al.,Lerobot: State-of-the-art machine learning for real-world robotics in pytorch,https : / / github . com / huggingface/lerobot, 2024

    2024

  19. [19]

    Mobile aloha: Learning bimanual mobile manipulation with low-cost whole-body teleoperation,

    Z. Fu, T. Z. Zhao, and C. Finn, “Mobile aloha: Learning bimanual mobile manipulation with low-cost whole-body teleoperation,” in Conference on Robot Learning (CoRL), 2024. arXiv:2401.02117 [cs.RO]

    Pith/arXiv arXiv 2024

  20. [20]

    J. Zhou, B. Liang, J. Huang, I. Zhang, and M. Tomizuka,Global- local interface for on-demand teleoperation, 2025. arXiv:2502. 09960 [cs.RO]

    2025

  21. [21]

    Iyer et al.,Open teach: A versatile teleoperation system for robotic manipulation, 2024

    A. Iyer et al.,Open teach: A versatile teleoperation system for robotic manipulation, 2024. arXiv:2403.07870 [cs.RO]

    Pith/arXiv arXiv 2024

  22. [22]

    A survey of robot learning from demonstration,

    B. D. Argall, S. Chernova, M. Veloso, and B. Browning, “A survey of robot learning from demonstration,”Robotics and Autonomous Systems, vol. 57, no. 5, pp. 469–483, 2009

    2009

  23. [23]

    Dexforce: Extracting force-informed actions from kinesthetic demonstrations for dexterous manipulation,

    C. Chen, Z. Yu, H. Choi, M. Cutkosky, and J. Bohg, “Dexforce: Extracting force-informed actions from kinesthetic demonstrations for dexterous manipulation,”IEEE Robotics and Automation Letters (RA-L), 2025. arXiv:2501.10356 [cs.RO]

    Pith/arXiv arXiv 2025

  24. [24]

    Multimodal and force-matched imitation learn- ing with a see-through visuotactile sensor,

    T. Ablett et al., “Multimodal and force-matched imitation learn- ing with a see-through visuotactile sensor,”IEEE Transactions on Robotics (T-RO), vol. 41, pp. 946–959, 2025. arXiv:2311.01248 [cs.RO]

    Pith/arXiv arXiv 2025

  25. [25]

    Zhang, C

    D. Zhang, C. Yuan, C. Wen, H. Zhang, J. Zhao, and Y . Gao, Kinedex: Learning tactile-informed visuomotor policies via kines- thetic teaching for dexterous manipulation, 2025. arXiv:2505 . 01974 [cs.RO]

    2025

  26. [26]

    Kinesoft: Learning proprioceptive manipulation policies with soft robot hands,

    U. Yoo, J. Francis, J. Oh, and J. Ichnowski, “Kinesoft: Learning proprioceptive manipulation policies with soft robot hands,” in Conference on Robot Learning (CoRL), 2025. arXiv:2503.01078 [cs.RO]

    Pith/arXiv arXiv 2025

  27. [27]

    Airexo-2: Scaling up generalizable robotic imita- tion learning with low-cost exoskeletons,

    H. Fang et al., “Airexo-2: Scaling up generalizable robotic imita- tion learning with low-cost exoskeletons,” inConference on Robot Learning (CoRL), 2025. arXiv:2503.03081 [cs.RO]

    Pith/arXiv arXiv 2025

  28. [28]

    Liang et al.,Universal manipulation exoskeleton: Learning com- pliant whole-body policies with real-time torque feedback, 2026

    L. Liang et al.,Universal manipulation exoskeleton: Learning com- pliant whole-body policies with real-time torque feedback, 2026. arXiv:2606.14218 [cs.RO]

    arXiv 2026

  29. [29]

    Zhong et al.,Nuexo: A wearable exoskeleton covering all upper limb rom for outdoor data collection and teleoperation of humanoid robots, 2025

    R. Zhong et al.,Nuexo: A wearable exoskeleton covering all upper limb rom for outdoor data collection and teleoperation of humanoid robots, 2025. arXiv:2503.10554 [cs.RO]

    Pith/arXiv arXiv 2025

  30. [30]

    Open-television: Teleoperation with immersive active visual feedback,

    X. Cheng, J. Li, S. Yang, G. Yang, and X. Wang, “Open-television: Teleoperation with immersive active visual feedback,” inConference on Robot Learning (CoRL), 2024. arXiv:2407.01512 [cs.RO]

    Pith/arXiv arXiv 2024

  31. [31]

    Ding et al.,Bunny-visionpro: Real-time bimanual dexterous teleoperation for imitation learning, 2024

    R. Ding et al.,Bunny-visionpro: Real-time bimanual dexterous teleoperation for imitation learning, 2024. arXiv:2407 . 03162 [cs.RO]

    2024

  32. [32]

    Y . Gao, H. Ma, and P. Zheng,Glovity: Learning dexterous contact- rich manipulation via spatial wrench feedback teleoperation system,

  33. [33]

    Dexcap: Scalable 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: Scalable and portable mocap data collection system for dexterous manipulation,” inRobotics: Science and Systems (RSS),

  34. [34]

    J. Koh, H. Jung, N. Kim, W. Ko, and C. Nam,Dex-mouse: A low- cost portable and universal interface with force feedback for data collection of dexterous robotic hands, 2026. arXiv:2604.15013 [cs.RO]

    Pith/arXiv arXiv 2026

  35. [35]

    Forcemimic: Force-centric imitation learning with force-motion capture system for contact-rich manipulation,

    W. Liu, J. Wang, Y . Wang, W. Wang, and C. Lu, “Forcemimic: Force-centric imitation learning with force-motion capture system for contact-rich manipulation,” inIEEE International Conference on Robotics and Automation (ICRA), 2025. arXiv:2410.07554 [cs.RO]

    Pith/arXiv arXiv 2025

  36. [36]

    A. Zhu, M. Zhu, B. J. Kim, Q. Wang, J. V . S. H. Ramos, and D. Hong,Midas hand: Modular low-impedance direct-drive anthro- pomorphic sensing hand, 2026. arXiv:2607.14487 [cs.RO]

    Pith/arXiv arXiv 2026

  37. [37]

    Zhang et al.,Mediapipe hands: On-device real-time hand tracking,

    F. Zhang et al.,Mediapipe hands: On-device real-time hand tracking,

  38. [38]

    Dinov2: Learning robust visual features without supervision,

    M. Oquab et al., “Dinov2: Learning robust visual features without supervision,”Transactions on Machine Learning Research (TMLR),

  39. [42]

    arXiv:2304.07193 [cs.CV]

    Pith/arXiv arXiv

  40. [2020]

    arXiv:2006.10214 [cs.CV]

    Pith/arXiv arXiv 2006

  41. [2024]

    arXiv:2403.07788 [cs.RO]

    Pith/arXiv arXiv

  42. [2025]

    arXiv:2510.09229 [cs.RO]

    arXiv