Pith. sign in

REVIEW 4 major objections 5 minor 52 references

RUKA: Rethinking the Design of Humanoid Hands with Learning

T0 review · 4 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read RUKA: a $1,300 tendon-driven humanoid hand that outperforms LEAP and Allegro on reachability, durability, and strength.

desk verdict A genuinely useful open-source hand whose controller claims currently rest on a self-referential measurement loop; the hardware is a contribution, the evaluation needs another pass. read the letter →

arxiv 2504.13165 v1 pith:6RT4ZPUD submitted 2025-04-17 cs.RO cs.AI

Anya Zorin , Irmak Guzey , Billy Yan , Aadhithya Iyer , Lisa Kondrich , Nikhil X. Bhattasali , Lerrel Pinto This is my paper · ORCID
classification cs.ROcs.AI
keywords tendon-drivenhandunderactuateddexterouslearnedcontrolmotion-captureglovedatateleoperationopen-sourceroboticshardware3D-printedrobotgrasptaxonomy
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper introduces RUKA, a 3D-printed, tendon-driven humanoid hand that costs under $1,300 and has 15 degrees of freedom driven by 11 motors in the forearm. Its central claim is that this low-cost, human-sized hand matches or beats popular direct-drive research hands on reachability, durability, and strength, and that learned controllers can handle the tendon nonlinearities that make such hands hard to control. The authors fit a commercial human motion-capture glove directly onto the robot hand, collect pose-and-command pairs by random motor walks, and train per-finger long short-term memory plus multi-layer perceptron models to map desired fingertip positions and joint angles to motor commands. If correct, this dissolves the usual trade-off between compactness, power, and affordability, making capable dexterous hands accessible to labs with a 3D printer.

What carries the argument

The load-bearing element is the physical and sensory loop around morphological fidelity. Because RUKA is shaped like a human hand, a commercial motion-capture glove can be strapped onto the robot and used as an external proprioceptor, reporting fingertip positions and per-finger keypoints without joint encoders. This enables autonomous data collection: a random walk over motor positions generates thousands of observed pose and motor-command pairs. Each finger gets a long short-term memory network that ingests the past 10 pose observations, and a multi-layer perceptron head that outputs the next motor positions, trained by mean squared error loss. The same glove then serves as the teleoperation input. The tendon-and-spring transmission, with braided line routed through low-friction tubes and extension springs for extension, is what lets the actuators live in the forearm, removing finger weight and enabling the reported strength values.

What would settle it

Mount joint encoders on RUKA or track it with an external optical motion-capture system while the hand runs the same random motor walks and controller evaluations, and compare the encoder angles to the glove-derived keypoints. If the glove misreads the coupled distal-proximal interphalangeal joints by more than the reported tolerances, such as more than the 3 millimeter transfer error claimed, then the learned controllers are fitting glove artifacts rather than true hand state. A second concrete check is to build a second RUKA with deliberately different tendon tension, run only the auto-calibration script, and measure fingertip errors on the human-validation set; the paper predicts a mean difference under 3 millimeters.

Watch

Extended reading notes

Core claim

RUKA is claimed to be a compact, affordable, capable tendon-driven hand: 3D-printed parts and off-the-shelf components, 15 underactuated degrees of freedom with 11 actuators, assembled in about 7 hours for under $1,300. Its morphology closely matches the average human hand, with five fingers and an opposable thumb; the distal and proximal interphalangeal joints of each non-thumb finger are coupled by one tendon, matching a human kinematic synergy. The paper reports that RUKA scores 10/10 on the Kapandji opposition test, reproduces 29 of 33 GRASP Taxonomy grasps, ran 20 hours continuously without precision loss, and outperformed the Allegro, LEAP, and Inmoov hands on pinch, payload, and slip tests, including a 2.74 N pinch force, 6.0 kg payload, and 33.02 N distal-plus-proximal interphalangeal slip force. To control the tendon-driven nonlinearities, the authors collect data by fitting a motion-capture glove to the hand and doing random walks over motor positions, then train separate controllers per finger; on human-validation poses, the thumb controller reaches mean errors around 0.83, 0.60, and 0.54 centimeters across the three axes, and a newly assembled hand replays poses within 3 millimeters.

Load-bearing premise

The whole control pipeline assumes that the commercial motion-capture glove, designed for human hands, returns accurate and stable fingertip and joint-angle readings when attached to a 3D-printed robot hand with nearby motors and moving tendons, and no independent joint encoder or external motion-capture check is reported.

Editorial extensions

If this is right

  • A lab with a 3D printer and roughly $1,300 in parts can build a five-fingered, human-sized hand that reports higher pinch, payload, and slip forces than the $2,000 LEAP and $15,000 Allegro hands in the paper's tests.
  • Because the controllers are trained from glove data, the same data-collection loop can be rerun after repairs or on a newly built hand, with only the auto-calibration script needing to run again.
  • Teleoperation at 25 Hz with glove input becomes feasible without joint encoders, which directly supports collecting human demonstrations for imitation learning; the paper demonstrates this with residual policies trained from human videos on cube flipping and bread pick-and-drop.
  • The 20-hour continuous runtime and sub-20-minute repairs address the overheating and repairability issues that typically limit direct-drive research hands.
  • The morphological accuracy claim implies that human hand data, such as glove poses and retargeted trajectories, transfers to RUKA with less manual retargeting than for non-anthropomorphic hands.

Reading between the lines

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

  • If the glove-as-sensor trick is as reliable as reported, the same recipe could extend beyond hands: any human-morphology robot body part, such as arms, legs, or feet, could use human wearable motion capture as ground truth, bypassing joint encoders in other tendon-driven designs.
  • The comparison table suggests the strength advantage comes mainly from removing actuator mass from the fingers; extending this logic, moving even more actuators proximally, into the forearm or torso, could further improve payload at the cost of more complex tendon routing.
  • The human-validation results hint that the learned controller generalizes across pose distributions better than nearest-neighbor retrieval, but the absence of tactile sensing suggests that adding simple joint encoders or inertial sensors could push errors below the reported centimeters and unlock finer precision tasks.
  • A direct test not in the paper: train the same controllers from an optical motion-capture source instead of the magnetic glove and compare human-validation errors; this would isolate whether the glove's sensors or the learning architecture is the bottleneck.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. RUKA is a 3D-printed, tendon-driven humanoid hand with 15 degrees of freedom driven by 11 actuators, assembled from off-the-shelf components for under $1,300. The paper's central claims are that (i) the hardware is compact, affordable, durable, and strong, outperforming popular hands such as LEAP, Allegro, and Inmoov on reachability, durability, and strength, and (ii) data-driven joint-to-actuator and fingertip-to-actuator controllers, trained from MANUS glove data collected on the robot, enable accurate pose replay and teleoperation. The manuscript includes hardware evaluations (Kapandji test, range-of-motion sampling, 33-grasp taxonomy, 20-hour endurance test, pinch/payload/slip tests), controller comparisons (LSTM+MLP versus MLP, k-NN, and search-based baselines), a transfer-to-a-new-hand experiment, teleoperation demonstrations, and HuDOR policy-learning rollouts. The design, code, and MuJoCo model are released open-source.

Significance. If the claims hold, this is a valuable open-source hardware contribution and a practical demonstration of learning-based control for tendon-driven hands. The design rationale is clear, the cost/DOF/DOA comparison in Table I is informative, and the strength numbers in Table III are concrete and reproducible in principle. The transfer-to-a-new-hand experiment, with a reported average difference of 3 mm, is a genuinely useful generalization test. However, the controller evaluation is built on the same MANUS glove used for training, with no independent ground-truth measurement, which weakens the central 'capable' claim and the absolute accuracy numbers in Table IV. The durability comparison and the grasp-count reporting also need tightening. The open-source release and detailed assembly instructions are strengths that make the suggested verification experiments feasible.

major comments (4)
  1. [V-A, V-B, Fig. 8] Controller training and all reported evaluation metrics use the same MANUS glove as ground truth. In Section V-A, data collection records glove keypoints and fingertip positions as the labels for motor positions; in Section V-B, both the Robot Validation targets and the replayed measurements are glove keypoints, and the Human Validation set is evaluated by replaying human glove data through the robot and comparing robot-glove keypoints with human-glove keypoints. Any systematic bias or distortion of the magnetic tracking when the glove is mounted on the 3D-printed hand near moving Dynamixel motors cancels between training and evaluation, so the sub-centimeter errors in Table IV do not establish absolute accuracy. The Limitations section (Section VII) names cost and lack of tactile sensing but not this missing ground-truth verification. Since the learned controllers are the paper's enabling contribution, please either validate the glove against joint encoders or external motion capture for at least a pose grid, or substantially soften the controller accuracy claims.
  2. [IV-B] Section IV-B reports that RUKA ran continuously for 20 hours and concludes that it 'outperforms both Allegro and LEAP hands,' but no comparison hand is run under the same protocol. The data shown in Fig. 5 are motor temperatures of RUKA only. Without a shared baseline, including the same actuation pattern, load, duration, and a defined failure criterion for the other hands, the 20-hour run supports an endurance demonstration, not a comparative durability claim. Please either run the comparison hands under matched conditions or rephrase the result as an absolute endurance result.
  3. [IV-A3, Fig. 7] The text in Section IV-A3 says RUKA 'successfully reproduces 29 out of 33 human hand grasps,' but the caption of Fig. 7 marks red grasps as not reached and yellow grasps as partially reached or unstable under perturbation. Counting partially reached or unstable poses as successes conflates the categories and inflates the reachability result. Please report the strict success count, the partial/unstable count, and the explicit criterion for each of the 33 poses in the GRASP Taxonomy, and reconcile the text with the figure caption.
  4. [IV-C2, IV-C3] The payload and slip tests define failure as 'joint angle error exceeds 15 degrees,' but the manuscript does not state how joint angles are measured in these tests. If they are obtained from the MANUS glove, the strength comparisons inherit the unverified glove bias discussed above; if they are measured by another method, that method should be described. Please specify the measurement instrument and, if it is the glove, validate it independently or report the strength results with the appropriate caveat.
minor comments (5)
  1. [III-B] The word 'mimicing' should be spelled 'mimicking'.
  2. [IV-D] In the phrase 'a 64.2% 1 improvement,' the footnote marker interrupts the text; place it after the percentage or at the end of the sentence.
  3. [Throughout] Several instances of missing spaces occur in the text, such as 'RUKAhas' and 'RUKAis'; a formatting pass would improve readability.
  4. [Table II] The CMC row lists the range as '190°−' with a dash where the human range should be; clarify whether this is a typo or whether no human comparator is available.
  5. [V-A, VI-A] Data collection is reported at 15 Hz and teleoperation at 25 Hz; please clarify how the 25 Hz controller operates when the training data were recorded at 15 Hz, or state that interpolation is used.

Circularity Check

1 steps flagged · score 4.0 of 10

Controller error metrics reduce to MANUS-glove self-consistency; hardware performance claims remain externally grounded.

  1. other [Section V-B2 (Controller Learning, Experiments), evaluation protocol; also Table IV]
    "During evaluation for both datasets, we attach the glove to the robot and replay the recorded keypoints by moving the robot motors to the positions predicted by the controllers. We then record the robot's keypoints and measure accuracy by comparing the reproduced fingertip positions against the originally saved ones."

    Training data (V-A) are (MANUS glove fingertip/joint positions, motor commands) pairs with the glove on RUKA and no joint encoders or external mocap. Evaluation (V-B2) uses the same MANUS glove to record both the saved target keypoints and the replayed robot keypoints. With glove reading G(p)=h(p)+b, where b is any robot-specific bias (motor magnetic interference, human-keypoint-model mismatch), Robot Validation error is |G(p_pred)-G(p_target)|=|h(p_pred)+b - h(p_target)-b|, so b cancels. Human Validation still measures the robot's response with the same glove, so robot-specific glove bias is invisible. Hence Table IV's errors are glove self-consistency, not physical fingertip accuracy; the controller evidence reduces to fit-and-replay of the same sensor.

full rationale

The paper's main hardware claims—cost, 15-DOF tendon-driven design, reachability (Kapandji, GRASP taxonomy), durability (20-hour run), and strength (pinch/payload/slip with scales)—are evaluated against external physical measurements and do not depend on the MANUS glove. The learned controllers, however, are trained and evaluated entirely in MANUS-glove coordinates: the glove defines both the target 'fingertip/joint positions' and the success metric. This creates a common-mode sensor bias in the reported controller errors, so the controller evaluation is partially self-referential. No load-bearing self-citation chain or imported uniqueness theorem appears; HuDOR and other same-lab works are used as applications or baselines, not to justify the core derivation. The Limitations section acknowledges reliance on the Manus glove only as a cost barrier and asserts it 'provides accurate data' without independent robot-side validation, which is exactly the missing external ground truth. Overall, the circularity is confined to the controller accuracy metrics; the hardware contribution retains independent empirical content.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central hardware and control claims rest on the reliability of glove-based measurement, the sufficiency of underactuated kinematics, and fair benchmarking against other hands. No new physical entities are introduced; the fitted quantities are per-hand motor calibration ranges and learned network weights.

free parameters (3)
  • Per-build motor limit calibration = not reported numerically
    The auto-calibration script (Section V-B2) binary-searches maximum motor ranges to account for tendon tensioning differences; controllers depend on these per-hand fitted limits. The paper reports only cross-run variation (0.5 degrees), not the fitted values.
  • Data collection sampling counts and random-walk steps = 500 thumb, 300 other fingers, 100 steps per walk
    These counts (Section V-A) are chosen by the authors without sensitivity analysis; they define the training set and therefore the accuracy of the learned inverse models.
  • LSTM/MLP hyperparameters = not specified
    The controller architecture (Section V-B1) uses a 10-step LSTM and an MLP head with MSE loss and AdamW, but hidden dimensions, layer counts, batch size, and learning rate are not reported, making exact reproduction impossible without contacting the authors.
assumptions (4)
  • domain assumption MANUS glove keypoints remain accurate when attached to a 3D-printed robot hand with motors, metal dowels, and moving tendons, and the derived joint angles are a faithful proxy for true joint angles.
    Section V-A collects all training labels from the glove without independent verification against encoders or external motion capture; corrupted glove data would propagate into every learned controller and evaluation.
  • domain assumption A hand that matches human size, finger count, and joint ranges can directly use human motion-capture data (and gloves) without retargeting.
    Section III-A1 motivates the anthropomorphic design by claiming it minimizes the need for retargeting, and Section V uses this to justify glove-based supervision. This is plausible but not quantitatively validated.
  • domain assumption Underactuated 15-DOF, 11-actuator kinematics with no active MCP abduction can reproduce the 33 GRASP taxonomy grasps well enough for the paper's claims.
    Section III-B and IV-A3 rely on underactuation being functionally sufficient; Fig. 7 itself marks some grasps as partially reached or unstable, qualifying this assumption.
  • ad hoc to paper The comparison hands (LEAP, Allegro, Inmoov) were tested under equivalent conditions, and their known durability limitations justify RUKA's superiority without running them for 20 hours.
    Section IV-B asserts RUKA outperforms Allegro and LEAP in durability based only on RUKA's 20-hour run; no endurance data for the baselines under the same protocol is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RUKA: Rethinking the Design of Humanoid Hands with Learning." pith.science (2026). https://pith.science/paper/6RT4ZPUD

@misc{pith2026250413165,
  author       = {Pith},
  title        = {Pith review of: RUKA: Rethinking the Design of Humanoid Hands with Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6RT4ZPUD}},
  note         = {Machine review of arXiv:2504.13165}
}
read the original abstract

Dexterous manipulation is a fundamental capability for robotic systems, yet progress has been limited by hardware trade-offs between precision, compactness, strength, and affordability. Existing control methods impose compromises on hand designs and applications. However, learning-based approaches present opportunities to rethink these trade-offs, particularly to address challenges with tendon-driven actuation and low-cost materials. This work presents RUKA, a tendon-driven humanoid hand that is compact, affordable, and capable. Made from 3D-printed parts and off-the-shelf components, RUKA has 5 fingers with 15 underactuated degrees of freedom enabling diverse human-like grasps. Its tendon-driven actuation allows powerful grasping in a compact, human-sized form factor. To address control challenges, we learn joint-to-actuator and fingertip-to-actuator models from motion-capture data collected by the MANUS glove, leveraging the hand's morphological accuracy. Extensive evaluations demonstrate RUKA's superior reachability, durability, and strength compared to other robotic hands. Teleoperation tasks further showcase RUKA's dexterous movements. The open-source design and assembly instructions of RUKA, code, and data are available at https://ruka-hand.github.io/.

Figures

Figures reproduced from arXiv: 2504.13165 by the authors.

Figure 1
Figure 1. RUKA is a tendon-driven humanoid hand that is simple, affordable, and capable. Its size and morphology closely match those of a human hand, enabling it to perform diverse human-like power, precision, and fine-grained grasps. increases the hand’s size and weight [44, 1]. To prioritize compactness and strength, other hands adopt a tendon-driven design with actuators outside the hand [43, 11], but this introduces nonli… view at source ↗
Figure 2
Figure 2. (A) A Venn diagram of a variety of robotic hands [1, 44, 43, 13, 24, 11] demonstrates RUKA’s unique combination of low cost, anthropomorphism and usability. (B) An illustration of the sizes of different hands that are commonly used by the robotics community. RUKA is designed to closely match the average human hand. and actuation technologies. However, we argue that learning￾based approaches present an opportunity to… view at source ↗
Figure 3
Figure 3. (A) Joints enable 15 degrees of freedom of RUKA labeled with their corresponding joint names. (B) The splay of the fingers allow for natural abduction-adduction movement without an active degree of freedom. (C) The MCP and PIP / DIP coupled tendons (light blue and dark blue respectively) are responsible for flexion, while the springs are responsible for extension. hand is underactuated, previous studies on robotic h… view at source ↗
Figures from the paper (7 more)
Figure 5
Figure 5. Figure 5: We run the hand continuously for 90 minutes, repeatedly doing [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: The experimental setup for the strength tests conducted on [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]
Figure 7
Figure 7. Figure 7: RUKA doing the 29 out of 33 Grasps from the GRASP Taxonomy [14], including a variety of power grasps, precision grasps, and intermediate grasps. Red grasps were not reached and yellow were partially reached or unstable to perturbations [PITH_FULL_IMAGE:figures/full_fi…
Figure 8
Figure 8. Figure 8: Keypoints received from the MANUS Haptic Gloves (left and right) and the controller architecture (center). Fingertip positions [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]
Figure 9
Figure 9. Figure 9: RUKA’s controller testing framework. The top row shows how we collect our test data on the powered-off RUKA hand and bottom row shows corresponding controller performance on the desired poses. positions, keypoints, joint angles, and both commanded and actual motor posi…
Figure 10
Figure 10. Figure 10: Teleoperated tasks made possible with RUKA. The top three rows showcase teleoperation of in-hand grasping tasks, while the remaining rows demonstrate arm-and-hand teleoperated tasks [PITH_FULL_IMAGE:figures/full_fig_p009_10.png]
Figure 11
Figure 11. Figure 11: Rollouts of tasks learned with HuDOR [19] on RUKA. For each task, the bottom row shows the open-loop rollout of the human trajectory, while the top row shows the final rollout after online finetuning, highlighting the importance of the training process. B. Policy Lear…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

52 extracted references · 24 canonical work pages

  1. [1]

    Allegro, 2010

    allegro. Allegro, 2010. URL https://www.allegrohand. com/ah-v4

    work page 2010

  2. [2]

    Holo-dex: Teaching dexterity with immersive mixed reality, 2022

    Sridhar Pandian Arunachalam, Irmak G ¨uzey, Soumith Chintala, and Lerrel Pinto. Holo-dex: Teaching dexterity with immersive mixed reality, 2022

    work page 2022

  3. [3]

    Bambu 3d printer, 2010

    bambux1c. Bambu 3d printer, 2010. URL https://us. store.bambulab.com/products/x1-carbon

    work page 2010

  4. [4]

    King, Daniele Moro, Kai-Hung Chang, Stelian Coros, and Nancy Pollard

    Dominik Bauer, Cornelia Bauer, Jonathan P. King, Daniele Moro, Kai-Hung Chang, Stelian Coros, and Nancy Pollard. Design and control of foam hands for dexterous manipulation.International Journal of Humanoid Robotics, 17(1), February 2020

    work page 2020

  5. [5]

    All the feels: A dexterous hand with large area sensing.arXiv preprint arXiv:2210.15658, 2022

    Raunaq Bhirangi, Abigail DeFranco, Jacob Adkins, Carmel Majidi, Abhinav Gupta, Tess Hellebrekers, and Vikash Kumar. All the feels: A dexterous hand with large area sensing.arXiv preprint arXiv:2210.15658, 2022

    arXiv 2022

  6. [6]

    Comparison between rigid and soft poly- articulated prosthetic hands in non-expert myo-electric users shows advantages of soft robotics.Scientific Re- ports, 11, 12 2021

    Patricia Capsi Morales, Cristina Piazza, Manuel Cata- lano, Giorgio Grioli, lisa schiavon, elena fiaschi, and Antonio Bicchi. Comparison between rigid and soft poly- articulated prosthetic hands in non-expert myo-electric users shows advantages of soft robotics.Scientific Re- ports, 11, 12 2021. doi: 10.1038/s41598-021-02562-y

    work page doi:10.1038/s41598-021-02562-y 2021

  7. [8]

    Karen Liu

    Yuanpei Chen, Chen Wang, Li Fei-Fei, and C. Karen Liu. Sequential dexterity: Chaining dexterous policies for long-horizon manipulation, 2023. URL https://arxiv. org/abs/2309.00987

    arXiv 2023

  8. [9]

    Object-centric dexterous manipulation from human motion data.arXiv preprint arXiv:2411.04005, 2024

    Yuanpei Chen, Chen Wang, Yaodong Yang, and C Karen Liu. Object-centric dexterous manipulation from human motion data.arXiv preprint arXiv:2411.04005, 2024

    arXiv 2024

Show all 52 references
  1. [10]

    Churchill, L

    E. Churchill, L. L. Laubach, J. T. Mcconville, and I. Tebbetts.Anthropometric source book. Volume 1: Anthropometry for designers. NASA, 1978. URL https://ntrs.nasa.gov/citations/19790003563

    1978

  2. [11]

    Clone hand, 2010

    clone. Clone hand, 2010. URL https://clonerobotics.com/ hand

    2010

  3. [12]

    Dexhand, 2010

    dexhand. Dexhand, 2010. URL https://www.dexhand. org/

    2010

  4. [13]

    Robot utility models: General policies for zero-shot deployment in new environments

    Haritheja Etukuru, Norihito Naka, Zijin Hu, Seung- jae Lee, Julian Mehu, Aaron Edsinger, Chris Pax- ton, Soumith Chintala, Lerrel Pinto, and Nur Muham- mad Mahi Shafiullah. Robot utility models: General policies for zero-shot deployment in new environments. arXiv preprint arXi...

    2024 arXiv

  5. [14]

    Dollar, and Danica Kragic

    Thomas Feix, Javier Romero, Heinz-Bodo Schmied- mayer, Aaron M. Dollar, and Danica Kragic. The grasp taxonomy of human grasp types.IEEE Transactions on Human-Machine Systems, 46(1):66–77, 2016. doi: 10.1109/THMS.2015.2470657

    2016

  6. [15]

    Filaflex, 2010

    filaflex. Filaflex, 2010. URL https://recreus.com/gb/ filaments/24-filaflex-foamy.html

    2010

  7. [16]

    Neural network and jacobian method for solving the inverse stat- ics of a cable-driven soft arm with nonconstant curvature

    Michele Giorelli, Federico Renda, Marcello Calisti, An- drea Arienti, Gabriele Ferri, and Cecilia Laschi. Neural network and jacobian method for solving the inverse stat- ics of a cable-driven soft arm with nonconstant curvature. IEEE Transactions on Robotics, 31(4):823–834, 2...

    2015

  8. [17]

    See to Touch: Learning Tactile Dexterity through Visual Incentives.arXiv e-prints arXiv:2309.12300, September 2023

    Irmak Guzey, Yinlong Dai, Ben Evans, Soumith Chin- tala, and Lerrel Pinto. See to Touch: Learning Tactile Dexterity through Visual Incentives.arXiv e-prints arXiv:2309.12300, September 2023

    2023 arXiv

  9. [18]

    Dexterity from touch: Self-supervised pre-training of tactile representations with robotic play, 2023

    Irmak Guzey, Ben Evans, Soumith Chintala, and Lerrel Pinto. Dexterity from touch: Self-supervised pre-training of tactile representations with robotic play, 2023

    2023

  10. [19]

    Bridging the human to robot dexterity gap through object-oriented rewards.arXiv preprint arXiv:2410.23289, 2024

    Irmak Guzey, Yinlong Dai, Georgy Savva, Raunaq Bhi- rangi, and Lerrel Pinto. Bridging the human to robot dexterity gap through object-oriented rewards.arXiv preprint arXiv:2410.23289, 2024

    2024 arXiv

  11. [20]

    Teach a robot to fish: Versatile imitation from one minute of demonstrations, 2023

    Siddhant Haldar, Jyothish Pari, Anant Rai, and Lerrel Pinto. Teach a robot to fish: Versatile imitation from one minute of demonstrations, 2023. URL https://arxiv.org/ abs/2303.01497

    2023 arXiv

  12. [21]

    Dex- treme: Transfer of agile in-hand manipulation from sim- ulation to reality.arXiv, 2022

    Ankur Handa, Arthur Allshire, Viktor Makoviychuk, Aleksei Petrenko, Ritvik Singh, Jingzhou Liu, Denys Makoviichuk, Karl Van Wyk, Alexander Zhurke- vich, Balakumar Sundaralingam, Yashraj Narang, Jean- Francois Lafleche, Dieter Fox, and Gavriel State. Dex- treme: Transfer of agi...

    2022

  13. [22]

    Long short- term memory.Neural Comput., 9(8):1735–1780, Novem- ber 1997

    Sepp Hochreiter and J ¨urgen Schmidhuber. Long short- term memory.Neural Comput., 9(8):1735–1780, Novem- ber 1997. ISSN 0899-7667. doi: 10.1162/neco.1997.9.8

    1997 doi

  14. [23]

    Hume, Harris Gellman, Harry McKellop, and Robert H

    Mary C. Hume, Harris Gellman, Harry McKellop, and Robert H. Brumfield. Functional range of motion of the joints of the hand.The Jour- nal of Hand Surgery, 15(2):240–243, 1990. ISSN 0363-5023. doi: https://doi.org/10.1016/0363-5023(90) 90102-W. URL https://www.sciencedirect.com...

    1990

  15. [24]

    Inmoov, 2010

    inmoov. Inmoov, 2010. URL https://inmoov.fr/ inmoov-hand/

    2010

  16. [25]

    OPEN TEACH: A Versatile Teleopera- tion System for Robotic Manipulation.arXiv e-prints arXiv:2403.07870, March 2024

    Aadhithya Iyer, Zhuoran Peng, Yinlong Dai, Irmak Guzey, Siddhant Haldar, Soumith Chintala, and Ler- rel Pinto. OPEN TEACH: A Versatile Teleopera- tion System for Robotic Manipulation.arXiv e-prints arXiv:2403.07870, March 2024

    2024 arXiv

  17. [26]

    Kinematic synergies of hand grasps: a comprehensive study on a large publicly available dataset.Journal of NeuroEngineering and Rehabilitation, 2019

    Atzori M Jarque-Bou NJ, Scano A and M ¨uller H. Kinematic synergies of hand grasps: a comprehensive study on a large publicly available dataset.Journal of NeuroEngineering and Rehabilitation, 2019

    2019

  18. [27]

    Robust anthropo- morphic robotic manipulation through biomimetic dis- tributed compliance, 2024

    Kai Junge and Josie Hughes. Robust anthropo- morphic robotic manipulation through biomimetic dis- tributed compliance, 2024. URL https://arxiv.org/abs/ 2404.05262

    2024 arXiv

  19. [28]

    Kapandji

    A. Kapandji. Cotation clinique de l’opposition et de la contre-opposition du pouce.Annales de Chirurgie de la Main, 5(1):67–73, 1986. ISSN 0753-9053. doi: https://doi.org/10.1016/S0753-9053(86) 80053-9. URL https://www.sciencedirect.com/science/ article/pii/S0753905386800539

    1986 doi

  20. [29]

    Fluid lubricated dexterous finger mechanism for human-like impact absorbing capability.IEEE Robotics and Automa- tion Letters, 4(4):3971–3978, 2019

    Yong-Jae Kim, Junsuk Yoon, and Young-Woo Sim. Fluid lubricated dexterous finger mechanism for human-like impact absorbing capability.IEEE Robotics and Automa- tion Letters, 4(4):3971–3978, 2019. doi: 10.1109/LRA. 2019.2929988

    2019

  21. [30]

    Optimal control with learned local models: Application to dexterous manipulation

    Vikash Kumar, Emanuel Todorov, and Sergey Levine. Optimal control with learned local models: Application to dexterous manipulation. In2016 IEEE International Conference on Robotics and Automation (ICRA), pages 378–383, 2016. doi: 10.1109/ICRA.2016.7487156

    2016

  22. [31]

    Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101, 2017

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101, 2017

    2017 arXiv

  23. [32]

    Eureka: Human- Level Reward Design via Coding Large Language Mod- els.arXiv e-prints arXiv:2310.12931, October 2023

    Yecheng Jason Ma, William Liang, Guanzhi Wang, De- An Huang, Osbert Bastani, Dinesh Jayaraman, Yuke Zhu, Linxi Fan, and Anima Anandkumar. Eureka: Human- Level Reward Design via Coding Large Language Mod- els.arXiv e-prints arXiv:2310.12931, October 2023

    arXiv

  24. [33]

    Manus haptic gloves, 2010

    manus. Manus haptic gloves, 2010. URL https://www. manus-meta.com/

    2010

  25. [34]

    Principal components analysis based control of a multi- dof underactuated prosthetic hand.Journal of Neuro- Engineering and Rehabilitation, 7:16 – 16, 2010

    Giulia Matrone, Christian Cipriani, Emanuele Lindo Secco, Giovanni Magenes, and Maria Chiara Carrozza. Principal components analysis based control of a multi- dof underactuated prosthetic hand.Journal of Neuro- Engineering and Rehabilitation, 7:16 – 16, 2010. URL https://api.s...

    2010

  26. [35]

    Learning dexterous in-hand manipulation, 2019

    OpenAI, Marcin Andrychowicz, Bowen Baker, Maciek Chociej, Rafal Jozefowicz, Bob McGrew, Jakub Pa- chocki, Arthur Petron, Matthias Plappert, Glenn Powell, Alex Ray, Jonas Schneider, Szymon Sidor, Josh Tobin, Peter Welinder, Lilian Weng, and Wojciech Zaremba. Learning dexterous ...

    2019

  27. [36]

    The surprising ef- fectiveness of representation learning for visual imitation, 2021

    Jyothish Pari, Nur Muhammad Shafiullah, Sridhar Pan- dian Arunachalam, and Lerrel Pinto. The surprising ef- fectiveness of representation learning for visual imitation, 2021

    2021

  28. [37]

    An open- source anthropomorphic robot hand system: Hri hand.HardwareX, 7:e00100, 2020

    Hyeonjun Park and Donghan Kim. An open- source anthropomorphic robot hand system: Hri hand.HardwareX, 7:e00100, 2020. ISSN 2468-

    2020

  29. [38]

    Patel, Minas V

    Vatsal V . Patel, Minas V . Liarokapis, and Aaron M. Dollar. Open robot hardware: Progress, benefits, chal- lenges, and best practices.IEEE Robotics & Automation Magazine, 30(3):123–148, 2023. doi: 10.1109/MRA. 2022.3225725

    2023

  30. [39]

    Rbo hand 3: A platform for soft dexterous manipulation.IEEE Transactions on Robotics, 38(6):3434–3449, 2022

    Steffen Puhlmann, Jason Harris, and Oliver Brock. Rbo hand 3: A platform for soft dexterous manipulation.IEEE Transactions on Robotics, 38(6):3434–3449, 2022. doi: 10.1109/TRO.2022.3156806

    2022

  31. [40]

    Dynamixel motors, 2010

    robotis. Dynamixel motors, 2010. URL https://www. robotis.us/dynamixel-xc330-m288-t/

    2010

  32. [41]

    Matthias Rolf and Jochen J. Steil. Efficient exploratory learning of inverse kinematics on a bionic elephant trunk. IEEE Transactions on Neural Networks and Learning Systems, 25(6):1147–1160, 2014. doi: 10.1109/TNNLS. 2013.2287890

    2014

  33. [42]

    King, Daniele Moro, Stelian Coros, and Nancy Pollard

    Cornelia Schlagenhauf, Dominik Bauer, Kai-Hung Chang, Jonathan P. King, Daniele Moro, Stelian Coros, and Nancy Pollard. Control of tendon-driven soft foam robot hands. In2018 IEEE-RAS 18th International Conference on Humanoid Robots (Humanoids), pages 1– 7, 2018. doi: 10.1109/...

    2018

  34. [43]

    Shadow hand, 2010

    shadow hand. Shadow hand, 2010. URL https://www. shadowrobot.com/dexterous-hand-series/. Shadow Hand

    2010

  35. [44]

    Leap hand: Low-cost, efficient, and anthropomorphic hand for robot learning.Robotics: Science and Systems (RSS), 2023

    Kenneth Shaw, Ananye Agarwal, and Deepak Pathak. Leap hand: Low-cost, efficient, and anthropomorphic hand for robot learning.Robotics: Science and Systems (RSS), 2023

    2023

  36. [45]

    Katzschmann

    Yasunori Toshimitsu, Benedek Forrai, Barnabas Gavin Cangan, Ulrich Steger, Manuel Knecht, Stefan Weirich, and Robert K. Katzschmann. Getting the ball rolling: Learning a dexterous policy for a biomimetic tendon- driven hand with rolling contact joints. In2023 IEEE-RAS 22nd Int...

    2023

  37. [46]

    A systematic approach to eval- uating and benchmarking robotic hands—the ffp in- dex.Robotics, 8(1), 2019

    Anand Vazhapilli Sureshbabu, Giorgio Metta, and Al- berto Parmiggiani. A systematic approach to eval- uating and benchmarking robotic hands—the ffp in- dex.Robotics, 8(1), 2019. ISSN 2218-6581. doi: 10.3390/robotics8010007. URL https://www.mdpi.com/ 2218-6581/8/1/7

    2019 doi

  38. [47]

    Karen Liu

    Chen Wang, Haochen Shi, Weizhuo Wang, Ruohan Zhang, Li Fei-Fei, and C. Karen Liu. DexCap: Scalable and Portable Mocap Data Collection System for Dex- terous Manipulation.arXiv e-prints arXiv:2403.07788, March 2024

    2024 arXiv

  39. [48]

    Hierarchical visual policy learning for long-horizon robot manipulation in densely cluttered scenes.arXiv preprint arXiv:2312.02697, 2023

    Hecheng Wang, Lizhe Qi, Bin Fang, and Yunquan Sun. Hierarchical visual policy learning for long-horizon robot manipulation in densely cluttered scenes.arXiv preprint arXiv:2312.02697, 2023

    2023 arXiv

  40. [49]

    Lessons from learn- ing to spin” pens”.arXiv preprint arXiv:2407.18902, 2024

    Jun Wang, Ying Yuan, Haichuan Che, Haozhi Qi, Yi Ma, Jitendra Malik, and Xiaolong Wang. Lessons from learn- ing to spin” pens”.arXiv preprint arXiv:2407.18902, 2024

    2024 arXiv

  41. [50]

    Design of a highly biomimetic anthropomorphic robotic hand towards ar- tificial limb regeneration

    Zhe Xu and Emanuel Todorov. Design of a highly biomimetic anthropomorphic robotic hand towards ar- tificial limb regeneration. In2016 IEEE International Conference on Robotics and Automation (ICRA), pages 3485–3492, 2016. doi: 10.1109/ICRA.2016.7487528

    2016

  42. [51]

    Robot synesthesia: In-hand ma- nipulation with visuotactile sensing

    Ying Yuan, Haichuan Che, Yuzhe Qin, Binghao Huang, Zhao-Heng Yin, Kang-Won Lee, Yi Wu, Soo-Chul Lim, and Xiaolong Wang. Robot synesthesia: In-hand ma- nipulation with visuotactile sensing. In2024 IEEE International Conference on Robotics and Automation (ICRA), pages 6558–6565....

    2024

  43. [672]

    URL https://www.sciencedirect.com/science/article/pii/ S2468067220300092

    doi: https://doi.org/10.1016/j.ohx.2020.e00100. URL https://www.sciencedirect.com/science/article/pii/ S2468067220300092

    2020 doi

  44. [1735]

    URL https://doi.org/10.1162/neco.1997.9.8.1735

    1997 doi

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.