Pith. sign in

REVIEW 3 major objections 5 minor 32 references

A hybrid hand puts soft material only at joints and keeps links rigid, matching rigid-hand repeatability while improving strength, endurance, and fragile-object teleoperation for under $600.

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 · grok-4.5

2026-07-14 22:26 UTC pith:S74ITKCK

load-bearing objection Solid open-hardware hand: hybrid joints + tendons deliver real strength/endurance and fragile-object gains at <$600; short fatigue horizon is the main caveat. the 3 major comments →

arxiv 2603.12120 v2 pith:S74ITKCK submitted 2026-03-12 cs.RO cs.AIcs.CV

CRAFT: A Tendon-Driven Hand with Hybrid Hard-Soft Compliance

Leo Lin , Shivansh Patel , Jay Moon , Svetlana Lazebnik , Unnat Jain This is my paper
classification cs.RO cs.AIcs.CV
keywords anthropomorphic robotic handtendon-driven actuationhybrid hard-soft compliancerolling-contact jointscontact-rich manipulationFeix grasp taxonomyopen-source robot handvision-based teleoperation
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.

Rigid robot hands break under contact; fully soft hands lose predictable shape and load capacity. This paper argues contact is not uniform: impacts concentrate at joints while links carry load. CRAFT therefore uses soft TPU only at the joints and rigid PLA for the links, with rolling-contact joint surfaces and a bidirectional linkage so the middle and tip joints flex together on a fixed path. Fifteen forearm-mounted motors drive the fingers through tendons, keeping the hand compact and light. Against a rigid baseline with the same motors, structural tests report higher pull-out strength, lower current during sustained holding, and comparable joint tracking error. Teleoperation improves success and speed on fragile and low-friction objects, and the hand completes all 33 grasps in the Feix taxonomy. The design is open-source under $600 with vision-based whole-arm teleoperation and simulation models, aimed at durable contact-rich data collection for learning.

Core claim

Localizing compliance at joints and rigidity at links, combined with rolling-contact geometry and PIP–DIP coupling, yields a tendon-driven anthropomorphic hand that preserves repeatable kinematics while passively absorbing contact. Relative to a rigid direct-drive baseline using the same motors, CRAFT reports higher finger pull-out strength, roughly half the current draw when holding a heavy load, similar sub-0.01 rad tracking error over an hour of cyclic grasping, higher teleoperation success on fragile and low-friction items, and full coverage of all 33 Feix grasp types, at under $600 and fully open-source.

What carries the argument

Hybrid hard-soft compliance with coupled rolling-contact joints: soft TPU at PIP/DIP (and thumb MP/IP) on rolling surfaces that distribute stress and constrain flexion to a repeatable path, plus a bidirectional linkage that equalizes those two joints, while rigid PLA links transmit tendon force and MCP/CMC joints use snap-fit interfaces for independent 2-DoF motion.

Load-bearing premise

One-hour cyclic grasp and static holding tests, plus short teleoperation trials, are enough to show that the soft rolling-contact joints will survive the high-cycle contact of large-scale robot learning data collection.

What would settle it

Run continuous grasp-release and heavy-hold protocols for many thousands of cycles or multi-day operation and check whether TPU rolling joints crack, uncouple, or lose the reported repeatability; early joint failure, or a teleop re-test with force sensing that shows no reduction in peak contact force versus the rigid baseline, would undermine the central durability and compliance claims.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Passive joint compliance can raise success on fragile and low-friction objects without active force control during teleoperation.
  • Tendon routing plus hybrid structure can deliver higher holding force and lower motor current than direct-drive rigid fingers with the same actuators.
  • Full Feix taxonomy coverage remains possible with soft joints if rolling contact and coupling keep motion paths fixed.
  • Open hardware under $600 with vision teleop and MuJoCo/URDF models can make contact-rich demonstration collection cheaper and less fragile.
  • Modular fingers and pop-out MCP joints can cut downtime after collisions during long data-collection runs.

Where Pith is reading between the lines

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

  • If soft-joint fatigue stays low over multi-day high-cycle use, hybrid hands may displace pure rigid platforms as the default for learning from contact-rich teleop.
  • Placing compliance only where impacts concentrate could transfer to other contact-rich limbs or tools without making entire structures soft.
  • Vision-only whole-arm teleop plus passive compliance may reduce reliance on expensive force sensing for demonstration collection.
  • Modeling rolling contacts as equality-constrained revolute joints may still leave a residual sim-to-real gap for force-sensitive policies that the paper does not fully close.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The manuscript introduces CRAFT, a fully 3D-printed, tendon-driven anthropomorphic hand (15 active DoF, ~800 g, under $600) that localizes soft TPU compliance at the joints and rigid PLA at the links, with rolling-contact PIP/DIP surfaces and a bidirectional PIP–DIP linkage for repeatable coupled flexion. Actuators sit in the forearm. Structural tests against a same-motor LEAP baseline report higher pull-out strength (15.29 N vs 8.67 N), comparable one-hour cyclic tracking error (<0.01 rad), and ~50% lower holding current under a 5 lb load. A five-object teleoperation study (10 trials per hand, arm fixed) shows higher success and lower completion times on fragile/low-friction items, and the hand is shown completing all 33 Feix grasps. URDF/MuJoCo models and vision-based whole-arm teleoperation are released with the open-source design.

Significance. If the reported structural and teleoperation gains hold under broader use, CRAFT is a practical, low-cost platform for contact-rich data collection and learning: hybrid compliance at the joints addresses a real failure mode of rigid open hands without the load and modeling penalties of fully soft designs, while the open release (hardware, teleop stack, sim models) lowers the barrier relative to Shadow, Allegro, or Leap V2. The same-motor LEAP comparison, photographic Feix coverage, and explicit kinematic comparison to RUKA/ORCA are concrete strengths that make the contribution usable by the community rather than purely conceptual.

major comments (3)
  1. [§V.A Structural Tests; §I, §VI] §V.A (Repeatability and Holding tests) and the framing in §I/§VI: the endurance claim rests on one-hour cyclic grasp–release and static 5 lb holding. For a hand motivated by large-scale robot-learning data collection, this horizon does not yet speak to multi-day high-cycle fatigue or abrasion of the TPU rolling-contact surfaces. The abstract and experiments do not over-claim multi-day life, but the introduction and conclusion repeatedly position CRAFT for long-running contact-rich collection; either a longer fatigue/abrasion protocol or an explicit, quantified limitation statement is needed so the durability claim is not over-read.
  2. [§V.B Teleoperation Tests] §V.B Teleoperation Tests: the arm is fixed at an optimal pre-grasp pose and only finger teleoperation is evaluated (10 trials per hand, different operators, 150 s cap). Success rates and times are reported without error bars, confidence intervals, or statistical tests. The design cleanly isolates hand compliance, but the claim that CRAFT “improves handling of fragile and low-friction items” in teleoperation is therefore limited to this constrained setting; a short note on statistical reliability and on how results may change under full whole-arm vision teleoperation (already demonstrated qualitatively in §V.D) would keep the claim proportionate.
  3. [Abstract; §III] Abstract vs. §III: the abstract states “Fifteen motors mounted on the fingers drive the hand through tendons,” while the body consistently places all actuators in the forearm/behind the wrist. This is a factual inconsistency in the paper’s primary claim sentence and should be corrected so the compact-form-factor argument is not misstated.
minor comments (5)
  1. [Fig. 7, Fig. 8] Fig. 7 and Fig. 8 captions and axis labels are hard to read in the manuscript text (heavy use of special characters / shading). Clarify joint-index legend and units so the <0.01 rad and current-draw comparisons are immediately verifiable.
  2. [§V.C Grasp Taxonomy Assessment] §V.C / Fig. 11: Feix coverage is photographic success under teleoperation. A one-sentence note that these are kinematic/static demonstrations (not force- or slip-quantified) would align the figure with how prior open hands report the taxonomy.
  3. [§IV Simulation] §IV Simulation: equality-constraint / revolute approximation of rolling contacts is reasonable for RL kinematics; briefly state whether tendon stretch or TPU hysteresis is omitted so users know the sim-to-real gap for force-sensitive policies.
  4. [§V.D, §II, §V.A] Typographical: “Whole Arm Teleopration” (§V.D heading); “Christophet al.” spacing; occasional “LEAPhand” concatenation. Standard copy-edit pass.
  5. [Table I; §V.A] Table I reports only strength; a single combined table of strength / mean tracking error / mean holding current would make the three structural claims easier to compare at a glance.

Circularity Check

0 steps flagged

No circularity: empirical hardware design validated against external baselines, with no fitted-as-prediction or self-definitional steps.

full rationale

CRAFT is a robotics hardware paper whose load-bearing claims are direct measurements (pull-out force 15.29 N vs LEAP 8.67 N, tracking error <0.01 rad over 1 h, ~50% lower holding current, teleop success/time on five objects, 33/33 Feix grasps) and kinematic comparisons to external hands (LEAP, RUKA, ORCA, Leap V2). The design premise that impacts concentrate at joints is an engineering hypothesis, not a quantity derived from or fitted to the reported metrics; the metrics are independent experimental outcomes under stated protocols (ArUco deflection, motor encoders, current draw, timed trials). There are no equations that reduce a claimed prediction to a fitted input, no uniqueness theorems imported from the authors, and no self-citations that carry the central results. Citations to prior hands supply baselines or related designs, not circular premises. The work is therefore self-contained against external benchmarks; score 0 is the correct outcome.

Axiom & Free-Parameter Ledger

2 free parameters · 3 axioms · 1 invented entities

As an empirical hardware paper the load-bearing content is design choices plus measurements, not free parameters fitted to a theory curve or newly postulated physical entities. The few free choices are engineering thresholds and material selections; the main axioms are standard domain assumptions about contact mechanics and 3-D-printed polymer behavior.

free parameters (2)
  • pull-out failure deflection threshold = 15°
    Authors define failure as fingertip deflection >15°; the numeric cutoff is chosen by hand and directly determines the reported strength numbers.
  • motor current thermal limit used for comparison = 600 mA
    600 mA motor limit is taken from the Dynamixel datasheet and used as the reference line in the holding test; not fitted but still a free engineering reference that frames the endurance claim.
axioms (3)
  • domain assumption Impacts and conforming contact concentrate at joints while links primarily transmit load, justifying soft material only at joints.
    Stated as the design premise in the Introduction and Abstract; never independently measured in the paper but used to motivate the entire hybrid architecture.
  • domain assumption Rolling-contact TPU surfaces distribute stress sufficiently to avoid the fatigue fracture typical of flexure joints under repeated bending.
    Invoked in Section III and Related Work with citation to Ahn et al.; short-term tests support it but long-term fatigue is assumed rather than proven.
  • domain assumption A linear map from operator joint range (HaMeR) to robot joint limits is sufficient for accurate vision-based teleoperation.
    Calibration procedure in Section IV; success of the user study rests on this retargeting assumption.
invented entities (1)
  • CRAFT hybrid hard-soft rolling-contact finger with bidirectional PIP-DIP TPU linkage no independent evidence
    purpose: Realize the hybrid compliance principle in a compact, fully 3-D-printable, tendon-driven anthropomorphic package.
    The specific combination of parts is new to this paper; independent evidence will exist only after third-party builds and long-term use, which the open-source release is intended to enable.

pith-pipeline@v1.1.0-grok45 · 16520 in / 2921 out tokens · 37854 ms · 2026-07-14T22:26:26.818138+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of CRAFT: A Tendon-Driven Hand with Hybrid Hard-Soft Compliance." pith.science (2026). https://pith.science/paper/S74ITKCK

@misc{pith2026260312120,
  author       = {Pith},
  title        = {Pith review of: CRAFT: A Tendon-Driven Hand with Hybrid Hard-Soft Compliance},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/S74ITKCK}},
  note         = {Machine review of arXiv:2603.12120}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

We introduce CRAFT hand, a tendon-driven anthropomorphic hand with hybrid hard-soft compliance for contact-rich manipulation. The design is based on a simple idea: contact is not uniform across the hand. Impacts concentrate at joints, while links carry most of the load. CRAFT places soft material at joints and keeps links rigid, and uses rollingcontact joint surfaces to keep flexion on repeatable motion paths. Fifteen motors mounted on the fingers drive the hand through tendons, keeping the form factor compact and the fingers light. In structural tests, CRAFT improves strength and endurance while maintaining comparable repeatability. In teleoperation, CRAFT improves handling of fragile and low-friction items, and the hand covers 33/33 grasps in the Feix taxonomy. The full design costs under $600 and will be released open-source with visionbased teleoperation and simulation integration. Project page: http://craft-hand.github.io/

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

32 extracted references · 6 linked inside Pith

  1. [1]

    Asymmetric rolling contact joint for enhanced payload capabilities.IEEE/ASME Transactions on Mechatronics, 29(1):202–213, 2023

    Jeongdo Ahn, Minho Hwang, Dukyoo Kong, Joonhwan Kim, and Dong-Soo Kwon. Asymmetric rolling contact joint for enhanced payload capabilities.IEEE/ASME Transactions on Mechatronics, 29(1):202–213, 2023

    2023

  2. [2]

    Learning dexterous in-hand manipula- tion.The International Journal of Robotics Research, 39 (1):3–20, 2020

    OpenAI: Marcin Andrychowicz, Bowen Baker, Maciek Chociej, Rafal Jozefowicz, Bob McGrew, Jakub Pa- chocki, Arthur Petron, Matthias Plappert, Glenn Powell, Alex Ray, et al. Learning dexterous in-hand manipula- tion.The International Journal of Robotics Research, 39 (1):3–20, 2020

    2020

  3. [3]

    Barretthand

    Barrett Technology. Barretthand. https://barrett.com/ barretthand. Accessed: 2025-12-24

    2025

  4. [4]

    All the feels: A dexterous hand with large area sensing.arXiv preprint arXiv:2210.15658, 2 (6), 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, 2 (6), 2022

    Pith/arXiv arXiv 2022

  5. [5]

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

    Tao Chen, Megha Tippur, Siyang Wu, Vikash Kumar, Ed- ward Adelson, and Pulkit Agrawal. Visual dexterity: In- hand reorientation of novel and complex object shapes. Science Robotics, 8(84):eadc9244, 2023

    2023

  6. [6]

    Christoph, Maximilian Eberlein, Filippos Katsimalis, Arturo Roberti, Aristotelis Sympetheros, Michel R

    Clemens C. Christoph, Maximilian Eberlein, Filippos Katsimalis, Arturo Roberti, Aristotelis Sympetheros, Michel R. V ogt, Davide Liconti, Chenyu Yang, Barn- abas Gavin Cangan, Ronan J. Hinchet, and Robert K. Katzschmann. Orca: An open-source, reliable, cost- effective, anthropomorphic robotic hand for uninter- rupted dexterous task learning. 2025. URL htt...

    arXiv 2025

  7. [7]

    A novel type of compliant and underactuated robotic hand for dexterous grasping.The International Journal of Robotics Re- search, 35(1-3):161–185, 2016

    Raphael Deimel and Oliver Brock. A novel type of compliant and underactuated robotic hand for dexterous grasping.The International Journal of Robotics Re- search, 35(1-3):161–185, 2016

    2016

  8. [8]

    The grasp taxonomy of human grasp types.IEEE Transactions on human-machine systems, 46(1):66–77, 2015

    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, 2015

    2015

  9. [9]

    Dlr-hit hand ii

    German Aerospace Center (DLR). Dlr-hit hand ii. https://www.dlr.de/en/rm/research/robotic-systems/ hands/dlr-hit-hand-ii. Accessed: 2025-12-24

    2025

  10. [10]

    Hand anthropometry of us army personnel

    Thomas M Greiner. Hand anthropometry of us army personnel. Technical report, 1991

    1991

  11. [11]

    Design of the utah/mit dextrous hand

    Steve Jacobsen, Edwin Iversen, D Knutti, R Johnson, and K Biggers. Design of the utah/mit dextrous hand. InProceedings. 1986 IEEE International Conference on Robotics and Automation, volume 3, pages 1520–1532. IEEE, 1986

    1986

  12. [12]

    Inmoov hand

    Ga ¨el Langevin. Inmoov hand. https://inmoov.fr/ inmoov-hand/. Accessed: 2026-01

    2026

  13. [13]

    Brl/pisa/iit softhand: a low-cost, 3d-printed, underactuated, tendon-driven hand with soft and adaptive synergies.IEEE Robotics and Automation Letters, 7(4): 8745–8751, 2022

    Haoran Li, Christopher J Ford, Matteo Bianchi, Manuel G Catalano, Efi Psomopoulou, and Nathan F Lepora. Brl/pisa/iit softhand: a low-cost, 3d-printed, underactuated, tendon-driven hand with soft and adaptive synergies.IEEE Robotics and Automation Letters, 7(4): 8745–8751, 2022

    2022

  14. [14]

    Multisensory five-finger dexterous hand: The dlr/hit hand ii

    Hong Liu, Ke Wu, Peter Meusel, Nikolaus Seitz, Gerd Hirzinger, MH Jin, YW Liu, SW Fan, T Lan, and ZP Chen. Multisensory five-finger dexterous hand: The dlr/hit hand ii. In2008 IEEE/RSJ international confer- ence on intelligent robots and systems, pages 3692–3697. IEEE, 2008

    2008

  15. [15]

    Yale openhand project: Optimizing open-source hand designs for ease of fab- rication and adoption.IEEE Robotics & Automation Magazine, 24(1):32–40, 2017

    Raymond Ma and Aaron Dollar. Yale openhand project: Optimizing open-source hand designs for ease of fab- rication and adoption.IEEE Robotics & Automation Magazine, 24(1):32–40, 2017

    2017

  16. [16]

    Isaac gym: High performance gpu-based physics simulation for robot learning.arXiv preprint arXiv:2108.10470, 2021

    Viktor Makoviychuk, Lukasz Wawrzyniak, Yunrong Guo, Michelle Lu, Kier Storey, Miles Macklin, David Hoeller, Nikita Rudin, Arthur Allshire, Ankur Handa, et al. Isaac gym: High performance gpu-based physics simulation for robot learning.arXiv preprint arXiv:2108.10470, 2021

    Pith/arXiv arXiv 2021

  17. [17]

    A framework for designing anthropomorphic soft hands through inter- action.arXiv preprint arXiv:2306.04784, 2023

    Pragna Mannam, Kenneth Shaw, Dominik Bauer, Jean Oh, Deepak Pathak, and Nancy Pollard. A framework for designing anthropomorphic soft hands through inter- action.arXiv preprint arXiv:2306.04784, 2023

    Pith/arXiv arXiv 2023

  18. [18]

    Passive Dynamic Walking.The In- ternational Journal of Robotics Research, 9(2):62–82,

    Tad McGeer. Passive Dynamic Walking.The In- ternational Journal of Robotics Research, 9(2):62–82,

  19. [19]

    URL http: //ijr.sagepub.com/content/9/2/62.abstract

    doi: 10.1177/027836499000900206. URL http: //ijr.sagepub.com/content/9/2/62.abstract

    doi:10.1177/027836499000900206

  20. [20]

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

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

    2020

  21. [21]

    Reconstructing hands in 3d with transformers

    Georgios Pavlakos, Dandan Shan, Ilija Radosavovic, Angjoo Kanazawa, David Fouhey, and Jitendra Malik. Reconstructing hands in 3d with transformers. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9826–9836, 2024

    2024

  22. [22]

    In-hand object rotation via rapid motor adaptation

    Haozhi Qi, Ashish Kumar, Roberto Calandra, Yi Ma, and Jitendra Malik. In-hand object rotation via rapid motor adaptation. InConference on Robot Learning, pages 1722–1732. PMLR, 2023

    2023

  23. [23]

    Frankmo- cap: Fast monocular 3d hand and body motion cap- ture by regression and integration.arXiv preprint arXiv:2008.08324, 2020

    Yu Rong, Takaaki Shiratori, and Hanbyul Joo. Frankmo- cap: Fast monocular 3d hand and body motion cap- ture by regression and integration.arXiv preprint arXiv:2008.08324, 2020

    Pith/arXiv arXiv 2008

  24. [24]

    Shadow dexterous hand

    Shadow Robot Company. Shadow dexterous hand. https: //www.shadowrobot.com

  25. [25]

    Demonstrating leap hand v2: Low-cost, easy-to-assemble, high-performance hand for robot learning.Robotics: Science and Systems (RSS), 2025

    Kenneth Shaw and Deepak Pathak. Demonstrating leap hand v2: Low-cost, easy-to-assemble, high-performance hand for robot learning.Robotics: Science and Systems (RSS), 2025

    2025

  26. [26]

    Leap hand: Low-cost, efficient, and anthropomor- phic hand for robot learning.arXiv preprint arXiv:2309.06440, 2023

    Kenneth Shaw, Ananye Agarwal, and Deepak Pathak. Leap hand: Low-cost, efficient, and anthropomor- phic hand for robot learning.arXiv preprint arXiv:2309.06440, 2023

    Pith/arXiv arXiv 2023

  27. [27]

    Robotic telekinesis: Learning a robotic hand imitator by watching humans on youtube, 2022

    Aravind Sivakumar, Kenneth Shaw, and Deepak Pathak. Robotic telekinesis: Learning a robotic hand imitator by watching humans on youtube, 2022. URL https://arxiv. org/abs/2202.10448

    Pith/arXiv arXiv 2022

  28. [28]

    Mujoco: A physics engine for model-based control

    Emanuel Todorov, Tom Erez, and Yuval Tassa. Mujoco: A physics engine for model-based control. In2012 IEEE/RSJ international conference on intelligent robots and systems, pages 5026–5033. IEEE, 2012

    2012

  29. [29]

    Getting the ball rolling: Learning a dexterous policy for a biomimetic tendon- driven hand with rolling contact joints

    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 International Conference on Humanoid Robots (Humanoids), pages 1–7. IEEE, 2023

    2023

  30. [30]

    Allegro hand

    Wonik Robotics. Allegro hand. https: //www.wonikrobotics.com/research-robot-hand. Accessed: 2025-12-24

    2025

  31. [31]

    How to create affordable, modular, light- weight, underactuated, compliant robot hands.National Technical University of Athens, Tech

    Agisilaos G Zisimatos, Minas V Liarokapis, Christo- foros I Mavrogiannis, George P Kontoudis, and Kostas J Kyriakopoulos. How to create affordable, modular, light- weight, underactuated, compliant robot hands.National Technical University of Athens, Tech. Rep, 2015

    2015

  32. [32]

    Bhattasali, and Lerrel Pinto

    Anya Zorin, Irmak Guzey, Billy Yan, Aadhithya Iyer, Lisa Kondrich, Nikhil X. Bhattasali, and Lerrel Pinto. Ruka: Rethinking the design of humanoid hands with learning.Robotics: Science and Systems (RSS), 2025

    2025