Working notes for Open-Source Humanoid Robotics Landscape (Oct 2026), compiled Oct 2, 2026. Not fact-checked line by line: where these notes and the report disagree, trust the report. See README.md.

Open-source dexterous hands, grippers, tactile sensors and hand-data capture: status as of 2026-10-02

Read this first: method and source limits

• Pages I could not read. The WebFetch proxy returned HTTP 429 (rate limit) on these, and I did not retry them by any other route:

  • robohorizon.com, ORCA Dexterity article (Mar 2026)
  • arxiv.org/html/2606.14561v1 and alphaxiv 2606.14561.md (I read the PDF instead)
  • orcahand.com
  • arxiv.org/abs/2603.26660 (I read the HTML v2 instead)
  • github.com/leap-hand/LEAP_Hand_API
  • github.com/pollen-robotics/AmazingHand
  • ruka-hand-v2.github.io
  • huggingface.co/blog/pollen-robotics/amazing-hand
  • leaphand.com
  • cnx-software.com, Sharpa Wave article
  • itcanthink.substack.com/p/robot-hands-are-getting-better

  Two more failures of other kinds:

  • roboticscenter.ai/store: refused because the shared WebFetch budget (800 calls/hour) ran out (HTTP 403).
  • humanoid.guide/product/linkerhand-l20: returned 404.

  Later, the WebSearch budget for the session (200 calls) also ran out.

• GitHub stats. github.com pages were rate-limited, so I could not take stars from them. Stars, forks, last push and latest default-branch commit date all come from GitHits package metadata, which is derived from the GitHub API. The latest-commit date is the Go-module pseudo-version date. All were fetched 2026-10-02, and the tool rounds counts of 1,000 or more (for example "2.4k"). Few of these repos have release tags; where there are none I give the latest commit date. PyPI dates are cross-checks.

• Source tags.

  • [WF] = opened with WebFetch today.
  • [GH] = README, docs or files read from the GitHub repo through GitHits.
  • [SR] = search-result title only, not opened.

• Prices. Each price is labelled as vendor, distributor, aggregator (humanoid.guide, robozaps) or cited in a paper. Aggregator and paper figures often disagree, so I give both.

• arXiv licences. The CC licence shown on arXiv HTML pages is the paper's licence. I did not treat it as the hardware licence.

Executive summary for a data, simulation or environment startup

1. Four open hands have real momentum in 2026:

   • ORCA: ETH spin-off, MIT licence, tactile and joint-encoder variants, LeRobot integration, release 2026-10-01.
   • Aero Hand Open: $314 BOM, MuJoCo model with spatial tendons, claims more than 400k cycles. Its CAD licence is non-commercial.
   • RUKA-v2: NYU, adds a 2-DoF wrist and abduction, about $1.5K.
   • The Hugging Face ecosystem: AmazingHand (2.4k stars) and the HOPEJr hand with a Hall-sensor teleop glove.

   LEAP is still the most-simulated research baseline, but it has slowed (last API push Oct 2025).

2. Closed hands now set the bar on durability and touch:

   • Sharpa Wave: about 1,000 taxels per fingertip, quoted at 1M–2.5M cycles, $50k from a distributor.
   • Mid-price Chinese hands: Inspire at about $5–6.5k, AgiBot OmniHand at $3,775.

   Against these, open hands win on hackability and licence, not on price or reliability.

3. Open tactile infrastructure is fragmented, and Meta's stack is largely archived. DIGIT design and interface, TACTO and Sparsh are all archived, and Sparsh is non-commercial. Tactile simulation is early: TacSL is "experimental" in Isaac Lab, and Taccel's release is unverified.

4. Hand-data capture is the weakest link:

   • Commercial gloves cost $1.5k–$7.3k or more.
   • Open gloves and exoskeletons are often non-commercial or gated (HOMIE, AirExo-2, GeoRT), or specific to one hand (DexUMI).
   • Open dexterous datasets are tiny next to gripper datasets: about 40–100 demos per task, versus FastUMI-100K's 100k trajectories (about 600 h).

1. Open-source dexterous hands

LEAP Hand family (v1, v2, v2 Advanced) — CMU's low-cost Dynamixel hand and the de facto research baseline

• Status: SLOWING.
  • Last push to leap-hand/LEAP_Hand_API was 2025-10-20.
  • The last hardware releases were LEAP v2 (RSS 2025) and v2 Advanced (IEEE Humanoids 2025).
  • I found no 2026 hardware.
  • The lead designer's site (fetched today) lists him as a final-year PhD student.
• Who. Kenneth Shaw (CMU Robotics Institute PhD student, designer); Deepak Pathak (CMU, advisor); Ananye Agarwal (v1 co-author). Shaw's site shows no company. Pathak co-founded Skild AI; this is prior knowledge, not verified today.
• Links.
  • GitHub: leap-hand/LEAP_Hand_API, leap-hand/LEAP_Hand_Sim
  • Web: leaphand.com, v2-adv.leaphand.com
  • Papers: arXiv 2309.06440 (v1, RSS 2023); RSS 2025 demo paper (v2)
• GitHub stats.
  • LEAP_Hand_API: 175 stars, 52 forks, last push 2025-10-20.
  • LEAP_Hand_Sim: 207 stars, 35 forks, last commit 2024-05-18.
• Licence.
  • LEAP_Hand_API: GitHub detects a non-standard ("other") licence, so check the terms before commercial use.
  • LEAP_Hand_Sim: MIT.
  • MuJoCo Menagerie model: MIT.
  • Hardware CAD licence: not verified.
• What it is.
  • v1 (2023).
    • 16 DoF: three fingers plus a thumb, 4 DoF each, with a "universal" MCP abduction mechanism that keeps all DoF at every flexion angle.
    • 16 Dynamixel XC330-M288-T servos, direct drive.
    • $2,000 BOM, under 4 h assembly, 595 g, about 30% larger than a human hand.
    • Python, ROS and C++ APIs.
  • v2 (RSS 2025 demo). $200, 8 DoF, hybrid rigid-soft, fully 3D-printed, assembled in under 2 h. Software includes MANO and Vision Pro retargeting, Manus-glove teleop, and URDFs for several simulators.
  • v2 Advanced (Humanoids 2025).
    • 21 DoF driven by 17 motors, including two powered palm articulations.
    • TPU+PLA multi-material print over a rigid internal structure; PIP and DIP coupled by one tendon.
    • About $3,000; can be built within a day.
• Adoption.
  • It is the comparison baseline in the ORCA, RUKA and RAPID papers.
  • RAPID Hand is pitched as a LEAP upgrade costing about $360 more.
  • Simulation coverage: MuJoCo Menagerie (leap_hand), a MuJoCo Playground manipulation environment, and dex-urdf.
  • XELA sells a uSkin kit "For LEAP Hand" [SR].
  • I found no units-sold or lab count.
• Benchmarks (v1 paper).
  • Pull-out grasp force 19.5 N, versus 8.5 N for Allegro.
  • Held 2 kg on one fingertip for 1 h with under 0.05 rad error.
  • Ran 1 h of 5 Hz grasp cycles. Under the same test, "Allegro begins to fail after 15 minutes".
  • Isaac Gym sim2real blind in-hand cube rotation.
  • The RUKA paper measured LEAP at 2.45 N pinch and 4.0 kg payload.
  • The ORCA paper found similar tracking accuracy and under 0.2 s latency, but LEAP motion was jerkier.
• Trade-offs.
  • Strengths: the most sim assets and prior work; direct drive with no tendon maintenance; easy to model.
  • Weaknesses: four fingers; oversized (v1); no tactile or joint sensors beyond the servos; Dynamixel cost.
  • I found no sign of wide uptake for v2 or v2 Adv.
  • Pick LEAP v1 to compare against RL and sim2real literature. Pick ORCA or RUKA for a human-sized five-finger hand.
• Sources.
  • [WF] ar5iv.labs.arxiv.org/html/2309.06440; roboticsproceedings.org/rss21/p132.html; v2-adv.leaphand.com; kennyshaw.net
  • [GH] Menagerie leap_hand/README; mujoco_playground manipulation listing; dex-urdf robots/hands listing
  • [SR] xelarobotics.com/products/for-leap-hand

ORCA Hand — ETH Zurich tendon-driven 17-DoF hand, now sold by ORCA Dexterity

• Status: ACTIVE.
  • orca_core tag v0.5.2 on 2026-10-01; PyPI orca-core 0.5.2 on the same day.
  • Platform paper arXiv 2606.14561 (June 2026).
  • A product-line launch was reported in March 2026 [SR].
• Who.
  • Soft Robotics Lab, ETH Zurich (PI Robert K. Katzschmann).
  • v1 authors include Clemens C. Christoph, Maximilian Eberlein, Filippos Katsimalis, Arturo Roberti, Aristotelis Sympetheros, Michel R. Vogt, Davide Liconti, Chenyu Yang, Barnabas G. Cangan and Ronan J. Hinchet.
  • Spin-off ORCA Dexterity, Inc.: a Delaware corporation with a Zurich team, per an aggregator.
  • 2026 platform paper: Francesco Capuano (Oxford), Maximilian Eberlein and Fabrice Bourquin (ETH), Clemens C. Christoph (Orca Dexterity).
  • Funding: not found.
• Links.
  • GitHub: orcahand/orca_core, orcahand/orcahand_description
  • Web: orcahand.com
  • Papers: arXiv 2504.04259 (v1), 2606.14561 (platform)
• GitHub stats. orca_core 597 stars, 156 forks, v0.5.2 (2026-10-01). orcahand_description stars were not retrieved.
• Licence. orca_core is MIT and orcahand_description is MIT; the platform paper says the stack is MIT. This is commercially reusable.
• What it is.
  • v1 (2025).
    • 17 DoF: 16 finger joints plus a 1-DoF belt-driven wrist (±60°). Fingers 2–5 have PIP, MCP and abduction; the thumb has IP, MCP, abduction and CMC.
    • One motor per DoF. Antagonistic 0.4 mm nylon tendons are routed through the joint centres.
    • "Poppable" joints that dislocate instead of breaking.
    • Automatic calibration of the tendon transmission ratio.
    • FSR fingertips giving binary touch.
    • Under 2,000 CHF in materials; under 8 h assembly.
  • v2 (2026).
    • orca_core ships configs for orcahand, -joint, -touch and -full, left and right. The full config uses joint encoders on all joints.
    • Motor options (paper): Dynamixel XC330-T288 or XC430-T240BB, or Feetech HL-3915 or HL-3930.
    • Optional Hall-effect fingertip force/torque sensors: 51 taxels on the thumb and pinky, 83 on the other fingers, 351 in total.
    • The API returns per-taxel 3-D positions and force vectors in palm or world frames.
  • Software. LeRobot-native; MuJoCo models and tasks (URDF and MJCF for v1 and v2); teleop through Meta Quest 3, MediaPipe or Manus (40–50 fps).
• Prices.

  • Platform paper: "ca. $3k". RUKA-v2 table: about $3.5K.

  • Aggregator robozaps (marked "verified 2026-07-31"):

    Variant	Actuators	Price
    Assembled base	Feetech	$3,500
    Assembled base	Dynamixel	$4,500
    Touch	Feetech	$6,100
    Touch	Dynamixel	$7,100
    "Lite" (self-build, coupled tendons)	—	about $900 BOM

    Weight 1.1 kg (base), about 1.25 kg (Touch).

  • A March 2026 headline said "3 open-source robotic hands, starting at $1,500" [SR].

  • Distributor SVRC lists Feetech STS3215 ×17, which conflicts with the paper.

• Adoption. No count of units or labs is published. 597 stars, a Discord community, and listings at distributors and aggregators.
• Benchmarks.
  • v1 durability: 2,250+ grasp cycles in 2.5 h without failure; an autonomous imitation-learning run of 7 h 17 min (about 2,000 grasps) with no hardware intervention.
  • v1 task results: pick-place IL about 90% (estimated from a figure); zero-shot sim2real ball reorientation after 1 h of IsaacGym training (4,096 environments).
  • v1 wear: tactile skin degraded after 2,000–4,000 cycles, and sensor wires snapped at 4,500–7,000 cycles.
  • v2: 5 h continuous (about 3,500 cycles) with no thermal cutoff and no actuator or transmission failure; learned in-hand reorientation succeeded 9/10.
• Trade-offs.
  • Strengths: human-sized five fingers plus a wrist; MIT licence; a live company; tactile and encoder options.
  • Weaknesses: tendons need tensioning and calibration; heavier than LEAP; durability is shown in hours and thousands of cycles, not millions; I found no public simulation model of its tactile sensors.
• Sources.
  • [WF] arxiv.org/html/2504.04259v2; arxiv.org/pdf/2606.14561; robozaps.com/products/orca-hand; roboticscenter.ai/hardware/orca-hand
  • [GH] orca_core README, tactile-frames.md, v2/orcahand-full-right/config.yaml, models listing; orcahand_description README and LICENSE
  • [SR] robohorizon headline (Mar 2026); Rowan Cheung post on X

RUKA and RUKA-v2 — NYU's human-sized tendon hand with learned controllers

• Status: ACTIVE. RUKA-v2 paper arXiv 2603.26660 (v2 dated 2026-03-30). The v1 repo's latest commit is 2025-08-11.

• Who. Lerrel Pinto's lab at NYU.

  • v1: Anya Zorin and Irmak Guzey (co-first authors), Billy Yan, Aadhithya Iyer, Lisa Kondrich, Nikhil X. Bhattasali, Lerrel Pinto.
  • v2: Xinqi (Lucas) Liu, Ruoxi Hu, Alejandro Ojeda Olarte, Zhuoran Chen, Kenny Ma, Charles Cheng Ji, Lerrel Pinto, Raunaq Bhirangi, Irmak Guzey.

• Links.

  • GitHub: ruka-hand/RUKA
  • Web: ruka-hand.github.io, ruka-hand-v2.github.io
  • Papers: arXiv 2504.13165 (v1), 2603.26660 (v2)
  • Data and controller weights: OSF

• GitHub stats. ruka-hand/RUKA 199 stars, 29 forks, latest commit 2025-08-11. I did not find a separate v2 code repo.

• Licence. The RUKA repo is MIT. The v2 hardware licence is not verified.

• What it is.

  • v1.
    • 15 DoF driven by 11 actuators: the thumb has 3 Dynamixel XM430-W210-T; each finger has 2 XL330-M288-T, with PIP and DIP coupled.
    • Actuators sit in the forearm; tendons are 200-lb braided line in PTFE tubes. Human-sized.
    • No joint encoders. Per-finger LSTM controllers are learned from MANUS-glove motion capture.
    • BOM under $1,300 (alternative motor configurations at $500 and $900). About 7 h assembly; most repairs under 20 min. MuJoCo model.
  • v2.
    • 20 total DoF, 16 actuated: 18 hand DoF including per-finger abduction (spring return), plus a 2-DoF wrist.
    • Optional AS5600 12-bit magnetic joint encoders and optional eFlesh fingertips.
    • About $1.5K.

• Benchmarks.

  • v1: pinch 2.74 N; payload 6.0 kg; ran 20 h continuously without a significant drop in precision; 29/33 grasps of the GRASP taxonomy; Kapandji 10/10.
  • v2 static holds: DIP-PIP 1.2 kg for 15 s; MCP 0.78 kg; thumb 0.835 kg for 20 s; wrist 1.215 kg.
  • v2 thermal: 5 h with no shutdown; the thumb settled at 46.97 °C.
  • v2 versus v1 in teleop: completion time down 51.3%, success up 21.2%.
  • v2 tasks: 13 teleoperated tasks; about 100 demos per learned task; average controller error 8.26°.

• Comparisons. RUKA-v2's Table 1 is the best cross-hand table found:

  Hand	Cost	DoF / actuated	Durability
  RUKA-v2	~$1.5K	20/16	>5 h
  ORCA	~$3.5K	17/17	>10K cycles
  Allegro V4	~$16K	16/16	—
  Wuji	~$5.5K	20/20	~40 min
  Sharpa Wave	~$50K	22/22	>2.5M cycles

• Trade-offs.

  • Strengths: the cheapest human-sized hand with a wrist; high payload.
  • Weaknesses: no encoders by default. The README says teleop works better with MANUS gloves because the controllers were trained on MANUS data, which ties you to a glove costing over $5k. Tendon tension varies from build to build.

• Sources.

  • [WF] arxiv.org/html/2504.13165v1; arxiv.org/html/2603.26660v2
  • [GH] RUKA README

Aero Hand Open — Chestnut Robotics' (formerly TetherIA) $314 tendon hand with 7 actuators

• Status: ACTIVE.
  • Latest commit 2026-09-15.
  • Paper arXiv 2608.28578 (Aug 2026).
  • Firmware 0.2.0 and a ROS2 release on 2026-02-25.
• Who. Chestnut Robotics, Inc. (formerly TetherIA, Inc.). Authors: Nan Wang, Mohit Yadav, Jonathan Wulff, Aidan Rosenbaum, Kezhou Chen, Yuvan Sharma, Xu Dong, Yiwei Tao.
• Links.
  • GitHub: Chestnut-Robotics/aero-hand-open; SDK tetheria/aero-open-sdk (PyPI 0.1.0.dev1)
  • Web: docs.tetheria.ai, shop.tetheria.ai
• GitHub stats. 976 stars, 154 forks, latest commit 2026-09-15.
• Licence: partially open.
  • Firmware and SDK: Apache-2.0.
  • CAD, STL, BOM and docs: CC BY-NC-SA 4.0 (non-commercial).
  • Commercial cloning or making parts needs a licence.
  • Units you buy may be integrated into commercial products.
• What it is.
  • 5 fingers, 16 joints, 7 actuated DoF: underactuated fingers and a thumb with 3-way coupling.
  • 7 Feetech HLS3606M coreless servos, about 0.59 N·m stall.
  • Kevlar/Vectran cables on 9 mm spools with music-wire return springs.
  • PLA print, about 13 h. ESP32-S3, USB-C, 6 V.
  • Weight 374 g (paper) or 389 g (README); 198×95×53.5 mm.
  • Proprioception from motor encoders only.
  • Simulation: MuJoCo model with 20 spatial tendons plus a 7-equation actuation map. Merged into MuJoCo Playground (2025-11-25) and Menagerie.
  • Teleop: ROS2 Humble; Manus (Oct 2025), webcam (Dec 2025), Apple Vision Pro (Jan 2026).
• Prices. Kit BOM $314 (README). Assembled $720 (shop, today), shipping in under 5 days.
• Benchmarks (self-reported).
  • About 12 N fingertip force (the shop page says about 10 N).
  • About 1.2 Hz open/close.
  • More than 400,000 full-actuation cycles without cable failure (90+ h).
  • All 33 GRASP-taxonomy grasps.
  • Zero-shot sim-to-real.
• Adoption. 976 stars; adopted into DeepMind's simulation assets. No sales numbers.
• Trade-offs.
  • Strengths: the cheapest credible hand, with a validated tendon simulation model.
  • Weaknesses: only 7 actuators; low force; non-commercial CAD. Release notes show recurring maintenance work: overheat protection, a derailment fix, and a homing change that broke backward compatibility.
• Sources.
  • [WF] arxiv.org/html/2608.28578v1; shop.tetheria.ai/products/aero-hand-open
  • [GH] README, RELEASE_NOTES.md; Menagerie listing
  • [SR] "TetherIA launches $314…"

AmazingHand — Pollen Robotics / Hugging Face four-finger hand for under $200

• Status: ACTIVE (borderline). Latest commit 2026-04-03; last push 2026-04-23.
• Who. Pollen Robotics (France), acquired by Hugging Face in April 2025 per Gigazine. The design is inspired by the ILDA hand.
• Links. GitHub pollen-robotics/AmazingHand; Hugging Face blog; Seeed wiki.
• GitHub stats. 2.4k stars, 280 forks.
• Licence. Code Apache-2.0; mechanical design CC BY 4.0 (per the Seeed wiki).
• What it is.
  • 4 fingers × 2 DoF (flexion/extension and abduction/adduction, from two parallel servos) = 8 DoF.
  • 8 Feetech SCS0009 micro servos.
  • About 400 g; printed rigid "bones" with TPU shells.
  • Control from Python over serial, or from an Arduino with a bus-servo driver.
  • BOM "less than $200" (Hugging Face blog title [SR]).
  • Force, durability and simulation model were not published in anything I read.
• Adoption. The most-starred open hand repo in this slice; Gitee mirrors; a MakerWorld listing.
• Trade-offs. Strengths: cheapest, easiest to build, permissive licences. Weaknesses: micro-servo forces, 4 fingers, no tactile sensing. Suited to education and human-robot-interaction work.
• Sources.
  • [WF] wiki.seeedstudio.com/hand_amazinghand; gigazine (HopeJR article, for the acquisition context)
  • [SR] Hugging Face blog title; Gigazine 2025-07-18 title

HOPEJr hand — The Robot Studio and Hugging Face hand and arm with a Hall-sensor teleop exoskeleton

• Status: ACTIVE. Latest commit 2026-05-04. LeRobot ships drivers for the hope_jr hand and arm and the homunculus_glove and exoskeleton.
• Who. Rob Knight (The Robot Studio, original design). Martino Russi (Hugging Face) did the 2025 redesign and the LeRobot integration.
• Links. GitHub TheRobotStudio/HOPEJr and nepyope/Project-Homunculus; LeRobot docs/source/hope_jr.mdx.
• GitHub stats. 831 stars, 101 forks.
• Licence. There is no LICENSE file at the repo root and GitHub detects none, so reuse rights are unclear.
• What it is.
  • 16 Feetech SCS0009 servos. The thumb has CMC, MCP, PIP and DIP. Each finger has a radial flexor, an ulnar flexor and coupled PIP/DIP.
  • Tendons are 0.33 mm, 63-lb line with PTFE guides and linearising spools; underactuated DIP; opposable thumb.
  • The exoskeleton and glove use SS49E Hall sensors, one per hand joint, on no-solder PCBs.
  • The full humanoid has 66 DoF and a target price of about $3,000 (Hugging Face, May 2025). The first few units were to ship by end-2025; actual shipping is not verified.
• Trade-offs. Strengths: the cheapest full arm, hand and teleop stack, LeRobot-native. Weaknesses: low force; unclear licence; no benchmarks.
• Sources.
  • [WF] gigazine.net/gsc_news/en/20250530-hugging-face-humanoid-robot
  • [GH] README, Arm/README.md, Arm/BOM.md, Arm/ChangeLog; LeRobot hope_jr.mdx

"DexHand" — three different projects share the name

• DexHand V1.0 (dexhand.org): DORMANT.
  • Designed by Rob Knight (The Robot Studio) and released as open source. Trent Shumay (IoT Design Shop) maintained the site, ROS 2 and BLE software.
  • The site's latest news is dated 2023-10-01.
  • Repos:
    • TheRobotStudio/V1.0-Dexhand: 577 stars, 101 forks, latest commit 2024-07-26.
    • iotdesignshop/dexhand_ros2_meta: 58 stars, 2024-06-30.
    • iotdesignshop/dexhand_description: 19 stars.
  • The ORCA authors call it "challenging to assemble" with no demonstrated real-world applicability. Specs not verified.
• BiDexHand (Northwestern, Zhengyang Kris Weng; arXiv 2504.14712, v2 Oct 2025): SLOWING.
  • Repo wengmister/BiDexHand: 257 stars, MIT, latest commit 2025-11-23.
  • 16 actuated DoF over 21 joints. N-configuration tendon routing, plus anti-parallelogram DIP–PIP linkages.
  • 2.14 N average peak fingertip force; lifts a 4.54 kg dumbbell; 33/33 GRASP grasps; Kapandji 9/11.
• DexRobot "DexHand021 Pro". A commercial 22-DoF hand [SR press-release title]. I could not verify any open hardware.
• Sources. [WF] dexhand.org; arxiv.org/html/2504.14712v1

InMoov hand — Gael Langevin's hobbyist tendon hand: SLOWING/DORMANT

• The tutorial page was last modified 2024-01-26. It recommends a successor, the "i2Hand", and has 458+ comment threads going back to 2014.
• 5 finger servos plus a wrist servo; braided 200-lb line.
• From the RUKA table: about $100, 14 DoF with 5 actuated, pinch 2.72 N, payload 3.2 kg.
• Licence is non-commercial CC BY-NC (prior knowledge; not stated on the page I fetched).
• Fine for education; not a manipulation-research platform.
• Sources. [WF] inmoov.fr/hand-and-forarm

Yale OpenHand — DORMANT

• Models: T, T42, O (commercialized by RightHand Robotics), Q, VF, W, M2, Stewart, Sphinx and F3.
• Licence CC BY-NC 3.0.
• grablab/openhand_node: 13 stars, MIT, latest commit 2022-02-10.
• Underactuated tendon and flexure grippers. Still useful as gripper designs, but not humanoid hands.
• Sources. [WF] eng.yale.edu/grablab/openhand

Pisa/IIT SoftHand (Natural Machine Motion Initiative) — DORMANT open CAD with a commercial descendant

• One motor drives 19 DoF through soft synergies.
• CAD is on SourceForge under CC BY 4.0.
• The commercial qb SoftHand2 is listed at $16,500 (aggregator).
• Robust power grasping; no in-hand dexterity.
• Sources. [WF] naturalmachinemotioninitiative.com/softhand; humanoid.guide (Wuji page list)

Brunel Hand (Open Bionics) and Will Cogley's designs — not verified this pass

• Will Cogley's Notion site is JavaScript-only and returned nothing.
• From prior knowledge: the Brunel Hand 2.0 is an older CC BY-SA design that is likely DORMANT or DISCONTINUED. Will Cogley publishes maker and YouTube biomimetic designs.

ISyHand (MPI-IS) — European open hand with an articulated palm

• Max Planck Institute for Intelligent Systems (Stuttgart) and the University of Augsburg: Benjamin A. Richardson, Felix Grüninger, Lukas Mack, Jörg Stückler, Katherine J. Kuchenbecker.
• Presented at Humanoids 2025.
• Dynamixel servos on the joints; about $1,300 in materials; 4 h assembly; Isaac simulation.
• In RL cube reorientation it outperforms comparable hands early in training and beats its own fixed-palm version.
• Repo names and stars not retrieved. Status: SLOWING or unverified.
• Sources. [WF] isyhand.is.mpg.de; arXiv 2509.26236

RAPID Hand (Sun Yat-sen University, UC Merced, CASIA; NeurIPS 2025)

• 20 DoF, fully actuated by 20 Dynamixel servos. 1.148 kg. About $3,500, of which the motors are about $1,800.
• 96 piezoresistive taxels per fingertip (480 total); Orbbec wrist RGB-D camera; sensor sync within 7 ms at 25 Hz; up to 7 N fingertip force.
• Results: rolling 50/50, translation 50/50, retrieval 24/50.
• Its comparison table lists Inspire at about $5,000.
• The paper says hardware will be released; release status unverified.
• Sources. [WF] arxiv.org/html/2506.07490v1

Closed comparators (prices with source type; date of fetch 2026-10-02 unless noted)

• Sharpa Wave.
  • $50,000 from distributor RoboticsCenter: "in stock", 4–6 weeks.
  • 22 active DoF; 0.8 kg; 20 N fingertip; 40 kg payload; more than 1,000 taxels per fingertip at 180 Hz; "2,500,000 cycles".
  • A December 2025 news item says rolling mass production began in October 2025, with a "one-million-cycle durability certification", 0.005 N sensitivity, and a CES 2026 Innovation Award.
  • Open URDF/USD/XML in sharpa-robotics/sharpa-urdf-usd-xml (24 stars, tag v4.0.0 2026-03-31).
  • MuJoCo Menagerie model added 2026-05-18.
  • The orca platform paper puts Sharpa at about $50k.
• Allegro (Wonik). V4 about $16K (RUKA-v2 paper; the LEAP paper says $16,000+). V5 Sense $13,000 (distributor). V5 got CE certification in May 2025 [SR]. Menagerie model available.
• Shadow Dexterous Hand. About $60k (orca 2026 paper), more than 100,000 CHF (ORCA v1 paper), or $100,000 (RUKA table, 22 DoF with 20 actuated). DEX-EE $60,000 (aggregator). Menagerie models for both.
• Inspire RH56DFX. About $5,000 (RAPID paper), $6,500 (aggregator), or "RH56 series" $10,500 (aggregator). 6 active DoF; 0.54 kg (aggregator).
• PSYONIC Ability Hand. $20,000 (aggregator). 6 actuators, 30 touch sensors (aggregator); 490 g and IP64 (vendor). Repo psyonicinc/ability-hand-api: 53 stars, MIT, last push 2026-05-18.
• Unitree Dex3-1 (vendor page). 3 fingers, 7 DoF, 33 tactile sensors, 710 g, 500 g maximum payload. No price listed.
• Unitree Dex5-1 (vendor page). 20 DoF (16 active plus 4), 1,100 g, 10 N fingertip, 3.5–4.5 kg payload. The Dex5-1P variant has 94 tactile sensors. No price listed.
• Wuji Hand. 20 active DoF, direct drive. About $5.5K (RUKA-v2) versus $16,000 (aggregator). "300,000+ grasp cycles" (aggregator) versus about 40 min thermal endurance (RUKA-v2).
• Robotera XHand1. 12 active DoF, fully actuated (DexUMI paper). $14,000 and 1.1 kg (aggregator).
• Tesollo DG-5F. 20 DoF, 1.76 kg, $21,000 (aggregator).
• Others listed by distributor or aggregator. AgiBot OmniHand $3,775 and OmniHand Pro $14,610 (distributor); ROBOTIS HX5-D20 $5,500 (aggregator); SCHUNK SVH $30,000 (aggregator); Agile Hand $42,000 (distributor).
• LinkerHand. Not verified (the aggregator page was 404).
• Sources. [WF] roboticscenter.ai/store/product/sharpa-wave-right; letsdatascience.com Sharpa news; humanoid.guide pages for wuji-hand, xhand1, dg-5f, ability-hand, rh56dfx and leap-hand; unitree.com/Dex3-1 and /Dex5-1; psyonic.io/ability-hand. [GH] Menagerie sharpa_wave README and CHANGELOG.

2. Grippers and handheld capture grippers

UMI (Universal Manipulation Interface) — handheld GoPro gripper for in-the-wild demos

• Status: ACTIVE (maintenance). Latest commit 2026-05-26.
• Who. Shuran Song's REAL lab (Columbia, now Stanford). Co-author list (Cheng Chi et al., with TRI) is from prior knowledge.
• Links. real-stanford/universal_manipulation_interface; arXiv 2402.10329.
• GitHub stats. 1.5k stars, 290 forks.
• Licence. MIT.
• What it is.
  • GoPro with a 155° fisheye lens and side mirrors for implicit stereo; IMU-aided ORB-SLAM3.
  • Soft fingers carrying fiducial markers to track gripper width.
  • 780 g; 310×175×210 mm; 80 mm stroke.
  • BOM $73 for the printed gripper plus $298 for the GoPro and accessories.
  • Menagerie has a umi_gripper model.
• Benchmarks.
  • More than 3× faster than teleop on cup arrangement.
  • Policies: cups 20/20, dynamic tossing 105/120, cloth folding 14/20, dish washing 14/20.
  • In the wild 43/60 (71.7%); cross-robot 18/20.
• Derivatives.
  • umi-on-legs: 548 stars; the default branch has not changed since 2024-08-29, but there was a push on 2026-07-20.
  • Also FastUMI, DexUMI, and exUMI (CoRL 2025).
• Trade-offs. The cheapest scalable demo device and hardware-agnostic, but parallel-jaw only. Its SLAM pipeline is fragile and depends on the GoPro and its firmware.
• Sources. [WF] arxiv.org/html/2402.10329v3. [GH] README (grep).

FastUMI / FastUMI-100K (Shanghai AI Lab)

• Status: ACTIVE. Repo MrKeee/FastUMI-100K: 58 stars, Apache-2.0, last push 2026-06-08.
• Who. Kehui Liu and Bin Zhao (Shanghai AI Lab), with Northwestern Polytechnical University, SJTU and others.
• What it is. A hardware-decoupled handheld gripper with plug-and-play fingertips. RealSense T265 for 6-DoF pose (200 Hz) plus a GoPro fisheye (60 Hz), synchronised at 20 Hz.
• Dataset. More than 100,000 trajectories (about 600 h) across 54 task categories, with 15,000 text annotations.
• Results. π0 fine-tuning reaches 73–100% on short-horizon tasks; transfer between xArm6 and Flexiv Rizon4 without fine-tuning.
• A "FastUMIPro" Hugging Face org exists [SR]; its commercial status is unverified.
• Sources. [WF] arxiv.org/html/2510.08022v1

Pollen PincOpen

• Status: ACTIVE. Latest commit 2026-06-16; 169 stars.
• Licence. CC BY-SA 4.0.
• What it is. A parallel gripper with a BOM of about €25, using a Feetech STS3215. It is a printable derivative of Reachy 2's "Pincette" gripper (BOM about €1,700) and fits the SO-ARM100.
• Community. A ROS2 driver from Christopher Newport University; a MuJoCo pull request; an ICRA 2026 workshop with Melexis 3D force sensing; LeKiwi integration.
• Sources. [GH] README

SO-101 gripper (The Robot Studio + Hugging Face)

• Status: ACTIVE. Last push 2026-09-23.
• TheRobotStudio/SO-ARM100: 7.6k stars, 708 forks, Apache-2.0.
• A 1-DoF jaw on an STS3215 servo. Kits from about 8 vendors: RobotEd, Robonine, PartaBot, ForgeMotion, Seeed, WowRobo, Autodiscovery and Phospho.
• Menagerie models: robotstudio_so101 and trs_so_arm100.
• The default entry gripper for LeRobot users.
• Sources. [GH] README (grep); Menagerie listing.

ALOHA gripper

• Status: DORMANT for the original repo: tonyzhaozh/aloha, 2.3k stars, MIT, latest commit 2024-04-19.
• Menagerie and Playground include ALOHA models.
• Commercial ALOHA 2 and Trossen kits continue (not verified).
• GELLO (arm teleop) is active: wuphilipp/gello_software, 551 stars, 2026-09-14.

3. Tactile sensors and tactile simulation

DIGIT (Meta, original) — DISCONTINUED as an open project

• Both repos are archived:
  • facebookresearch/digit-design: 216 stars.
  • facebookresearch/digit-interface: 93 stars; last release 0.2.1 on 2021-09-22.
• Commercial DIGIT through GelSight: unverified.

Digit 360 (Meta FAIR + GelSight) — multimodal fingertip

• Status: DORMANT (repo). Latest commit 2025-07-22.
• GitHub stats. facebookresearch/digit360: 264 stars, 21 forks.
• Licence. Custom ("other").
• Who. Mike Lambeta, Tingfan Wu, …, Jitendra Malik, Roberto Calandra.
• What it is.
  • About 8.3 million taxels (README).
  • 18+ sensing modalities, 7 µm features, 1 mN force (digit.ml).
  • On-device processing; ROS2 interface; meshes for Isaac Sim.
• Availability. Free units offered to researchers through a call for proposals with GelSight. Commercial price not found.
• Sources. [GH] README. [WF] digit.ml.

Sparsh (Meta FAIR) — tactile foundation encoders: DISCONTINUED (repo archived)

• facebookresearch/sparsh: 227 stars, archived; last push 2025-02-27.
• Licence. CC BY-NC 4.0 (non-commercial).
• Self-supervised training on 460k+ tactile images.
• TacBench: 6 tasks. Sparsh is reported +95.1% on average over task- and sensor-specific end-to-end models (Oct 2024).
• Sources. [WF] arxiv.org/abs/2410.24090. [GH] LICENSE.md.

ReSkin (CMU + Meta)

• Status: DORMANT. raunaqbhirangi/reskin_sensor: 134 stars, MIT, last push 2023-06-01.
• Magnetic skin whose electronics are separate from a replaceable elastomer.
• Sources. [WF] reskin.dev

AnySkin (NYU, CMU, Columbia, Meta)

• Status: DORMANT (library), but sold. anyskin 1.0.0 on 2024-10-21; repo 87 stars, MIT.
• Commercial units through WowRobo.
• 92% slip detection. First demonstration that learned policies transfer when the skin is swapped.
• The README says readings drift and must be re-zeroed.
• Sources. [WF] any-skin.github.io. [GH] README.

eFlesh (NYU)

• Status: SLOWING. notvenky/eFlesh: 384 stars, MIT, last push 2025-10-28.
• TPU-printed cut-cell microstructures with magnets and the ReSkin/AnySkin magnetometer boards.
• More than 90% average success on 4 visuo-tactile tasks.
• An optional fingertip on RUKA-v2.
• Sources. [GH] README

9DTact / DTact

• 9DTact: ACTIVE. linchangyi1/9DTact: 191 stars, MIT, last commit 2026-08-21. The README says commercial use is explicitly welcome.
  • Compact vision-based sensor for 3-D shape and 6-D force.
  • The README lists 16 "DTact series" papers from 2023 to 2026, including 2026 work on a low-cost tactile gripper, SpikingTac, ViHaTeleop and a heterogeneous tactile transformer.
• DTact: DORMANT. 21 stars, 2023.
• Sources. [GH] README

GelSlim 4.0 (MMint Lab, University of Michigan)

• Open design, code and data that build on GelSlim 3.0 (arXiv 2409.19770).
• Activity and stars unverified.

Evetac (TU Darmstadt group; IEEE T-RO)

• Event camera instead of RGB; 1,000 Hz online processing; senses vibration up to 498 Hz; shear and slip.
• Open design on sites.google.com/view/evetac.
• Activity unverified.

GelSight Mini (closed comparator)

• The vendor page lists no price ("request quote" or product sheet). About $500 at its 2022 launch is prior knowledge and unverified.
• SDK gelsightinc/gsrobotics: 206 stars, GPL-3.0, last push 2025-06-25.

TacSL in Isaac Lab (NVIDIA)

• Status: ACTIVE. Isaac Lab: 8.2k stars, BSD-3, pushed 2026-10-02.
• An experimental visuo-tactile sensor based on TacSL (Akinola, Xu, Carius, Fox, Narang; T-RO 2025) and the Taxim renderer.
• Outputs tactile RGB, depth, normal force and shear force.
• The examples are built on the GelSight R1.5, "a prototype sensor that is now discontinued".
• Sources. [GH] IsaacLab docs/experimental-features/visuo_tactile_sensor.rst, isaaclab_contrib README.

Taccel (Peking University, BIGAI, UCLA; NeurIPS 2025)

• Combines IPC and ABD solvers.
• Peg insertion: 915 FPS across 4,096 environments on an H100, about 18× real time.
• Dexterous hands: 12.67 FPS across 256 environments; 64× faster than SAPIEN-IPC.
• Sim-to-real classification: 86.5% on synthetic data, 70.94% on real.
• The paper says "will release"; release not verified.
• Sources. [WF] arxiv.org/html/2504.12908v2

Older tactile simulators

• TACTO: facebookresearch/tacto, 472 stars, archived.
• DiffTactile: Genesis-Embodied-AI/DiffTactile, 316 stars, latest commit 2024-03 (DORMANT).

4. Hand-data capture: gloves, exoskeletons and retargeting

DexUMI (Stanford, Columbia, J.P. Morgan AI Research, CMU, NVIDIA; CoRL 2025 Best Paper finalist)

• Status: ACTIVE. Latest commit 2026-04-18.
• GitHub stats. real-stanford/DexUMI: 266 stars, MIT.
• Who. Mengda Xu and Han Zhang (equal contribution), Yifan Hou, Zhenjia Xu, Linxi Fan, Manuela Veloso, Shuran Song.
• What it is.
  • A wearable exoskeleton designed per target hand: Inspire (6 active DoF) and XHand (12 active DoF).
  • Alps encoders; iPhone ARKit (Record3D) for wrist pose; OAK-1 150° camera.
  • FSR or electromagnetic tactile sensors.
  • SAM2 and ProPainter remove the human and exoskeleton from images and render the robot hand in their place.
• Results. 3.2× more efficient than teleop; 86% average success over four tasks.
• Authors' stated limitations. Each new hand needs its own exoskeleton. FSRs are "sensitive to attachment", and electromagnetic sensors drift after high pressure. Backlash and printed-part flex limit accuracy.
• Sources. [WF] arxiv.org/html/2505.21864v2. [GH] README.

DOGlove (Tsinghua TEA Lab, Shanghai Qi Zhi, Shanghai AI Lab; RSS 2025)

• Status: ACTIVE (borderline). Latest commit 2026-04-18.
• GitHub stats. TEA-Lab/DOGlove: 140 stars, MIT.
• What it is.
  • Under $600; under 6 h assembly; 550 g.
  • 21-DoF motion capture from 16 Alps encoders.
  • 5-DoF cable force feedback from Dynamixel XC/XL330 servos.
  • 5 fingertip linear resonant actuators for haptics.
  • Up to 120 Hz; ±1° after calibration.
• Benchmarks. With haptics and force feedback: 10/10 on in-hand rotation, versus 1/10 for AnyTeleop.
• Comparison. The paper puts Manus and SenseGlove each above $5,000.
• Sources. [WF] arxiv.org/html/2502.07730v1. [GH] README.

HOMIE / OpenHomie (Shanghai AI Lab + CUHK; RSS 2025)

• Status: ACTIVE. Last push 2026-09-08. A 2026-07 news item points to a HOMIE v2 RL policy in NVlabs' GR00T whole-body-control repo.
• GitHub stats. InternRobotics/OpenHomie: 619 stars.
• What it is. About $500 of hardware in total:
  • Isomorphic 7-DoF exoskeleton arms on XL330-M288-T servos.
  • Hall-sensor gloves with 15 DoF each.
  • A foot pedal for locomotion.
  • Targets the Unitree G1 with Dex3 hands.
• Licence. Hardware files require a form with a real name and institution, and the README says "strictly forbidden to use HOMIE for any commercial purposes".
• Sources. [WF] arxiv.org/html/2502.13013v2. [GH] README.

DexCap (Stanford)

• Status: DORMANT. Last push 2024-10-10.
• GitHub stats. j96w/DexCap: 421 stars, MIT.
• Rokoko electromagnetic-field (EMF) motion-capture gloves plus SLAM and 3-D observation; the DexIL learning method.
• Depends on a commercial glove.
• Sources. [WF] arxiv.org/abs/2403.07788. [GH] README.

AirExo-2 (AirExo team, Hongjie Fang et al.; SJTU per prior knowledge)

• Status: ACTIVE (light). Last push 2026-04-03; README news stops at June 2025.
• GitHub stats. 43 stars.
• Licence. CC BY-NC-SA 4.0 covering hardware, code and data.
• Arm exoskeletons for Flexiv Rizon 4 arms with Robotiq 2F-85 grippers, so not dexterous hands.
• Sources. [GH] README

Homunculus glove and exoskeleton (Hugging Face / nepyope)

• Status: ACTIVE. Latest commit 2026-09-23.
• GitHub stats. 46 stars; no licence detected.
• SS49E Hall sensors; the homunculus_glove teleop type in LeRobot.

Retargeting

• dex-retargeting. dexsuite: 1.2k stars, MIT, v0.5.0 on 2025-04-05, last push 2025-08-21. DORMANT by the 12-month rule, but stable and widely used.
• GeoRT (Meta). 231 stars, CC BY-NC 4.0, last push 2025-09-02. A neural retargeter that trains in 1–2 min.
• dex-urdf. 383 stars, MIT, last push 2025-08-17. URDFs for the Ability, Allegro, Barrett, D'Claw, Inspire, LEAP, Panda, Schunk and Shadow hands, but not ORCA, RUKA, Aero or XHand.
• VR teleop stacks.
  • unitreerobotics/xr_teleoperate: 1.6k stars, active 2026-09-07.
  • OpenTeleVision/TeleVision: 1.3k stars, 2024.
  • Dingry/BunnyVisionPro: 359 stars.

Closed gloves

• Rokoko Smartgloves. $1,495 (vendor). Hybrid IMU and EMF tracking, 7 sensors per glove, 100 fps.
• Manus Metagloves Pro. Quote only (vendor). EMF tracking, 25 DoF. The DOGlove paper puts it above $5,000. RUKA, ORCA, LEAP v2 and Aero all support Manus teleop.
• SenseGlove Nova 2. Listed at $7,299 excluding VAT and at €3,999 excluding VAT "for small quantities"; it is unclear whether these are per glove or per pair. 4 magnetic friction brakes up to 20 N; vibrotactile feedback on thumb and index.
• StretchSense. Not verified.
• Sources. [WF] rokoko.com/products/smartgloves; manus-meta.com/products/metagloves-pro; senseglove.com/product/nova-2

5. Also notable (brief)

• Tactile SoftHand-A (Bristol; arXiv 2406.12731): tactile, highly underactuated, 3D-printed hand. [SR] Status unverified.
• "Affordable, fully-actuated biomimetic hand" (arXiv 2510.16931): [SR] Unverified.
• HRI Hand (HardwareX; MrLacquer/HRI-hand-firmware): [SR] Likely DORMANT.
• CDF-Glove (arXiv 2603.05804, Mar 2026): cable-driven force-feedback glove. [SR] Open status unknown.
• Glovity (arXiv 2510.09229): wrench-feedback teleop. [SR]
• ViHaTeleop (arXiv 2608.16572, 2026): low-cost visual-haptic teleop built on 9DTact. [GH, 9DTact README]
• Universal Manipulation Exoskeleton (arXiv 2606.14218) and ExoGS (arXiv 2601.18629, real-to-sim-to-real from exoskeleton data). [SR]
• THETA (arXiv 2601.07768): hand-state estimation for teleop. [SR]
• Dexora (arXiv 2605.18722): open VLA for high-DoF bimanual hands, a consumer of hand data. [SR]
• DexWild (CMU; dexwild/dexwild): 46 stars, MIT, latest commit 2025-08-14. DORMANT.
• MuJoCo Menagerie: 4.1k stars, Apache-2.0. Became a pip package mujoco-menagerie in Sept 2026 (2026.9.4 on 2026-09-28). ACTIVE.
• MuJoCo Playground: 2.2k stars. Includes LEAP and Aero hand environments. ACTIVE.
• LeRobot: 27k stars, 0.6.1 on 2026-08-03. The integration hub for ORCA, HopeJR, PincOpen and SO-101. ACTIVE.

6. Comparison tables

Open hands (plus key closed comparators)

Tactile sensors and data-capture devices

7. Gaps and pain points (with evidence)

1. Tendon hands need constant maintenance, and durability numbers can't be compared.
   • ORCA v1 needs periodic manual re-tensioning; its skin wears out at 2–4k cycles and its wires at 4.5–7k.
   • Aero's README warns that tendons need regular tension adjustment and that overload damages servos. Its release notes add overheat cutoffs, a derailment fix and a homing change that broke compatibility.
   • RUKA's tension varies per build and needs a calibration script.
   • Open hands report durability in hours (5 h, 7 h, 20 h). Closed hands claim 300k to 2.5M cycles.
   • There is no shared durability or benchmark protocol; RUKA-v2's table mixes hours and cycles.
   • Opening: standardized real and simulated test suites for hands.
2. Weak proprioception means noisy labels.
   • RUKA has no encoders by default (learned controller; 8.26° average error in v2). Aero has motor encoders only. ORCA's joint encoders are an option.
   • Joint-state ground truth in data from these hands is therefore approximate, and sim2real relies on models of tendon coupling.
3. Simulation assets are patchy.
   • Menagerie covers LEAP, Allegro, Shadow and DEX-EE, Sharpa, Aero, UMI and SO-101.
   • Not covered: ORCA (which ships its own MJCF), RUKA, Inspire, XHand, Wuji, AmazingHand and the HOPEJr hand.
   • dex-urdf and dex-retargeting have been idle since Aug 2025.
   • Tendon and soft-skin dynamics are rarely modelled; Aero is the exception, with 20 spatial tendons.
   • Opening: maintained, system-identified sim twins of the popular open hands, including their tactile layers.
4. Tactile is fragmented and its sim support is weak.
   • Sensing types in use: binary FSRs, Hall-effect taxels (ORCA), magnetometer skins (AnySkin, eFlesh), piezoresistive arrays (RAPID) and gel cameras (GelSight-type).
   • Meta's open stack is archived (DIGIT, TACTO, Sparsh), and Sparsh is non-commercial.
   • Isaac Lab's TacSL is experimental and built around a discontinued GelSight prototype. Taccel's release is unverified.
   • Reliability is reported as poor: DexUMI found FSRs attachment-sensitive and electromagnetic sensors drifting; AnySkin needs re-zeroing; ORCA's skin wears out.
   • Opening: calibrated simulation models and datasets for the sensors that open hands actually ship.
5. Human-data capture is costly, closed or restricted.
   • Gloves: Manus is quote-only (above $5k), SenseGlove $7.3k, Rokoko $1.5k.
   • RUKA's controllers run best with MANUS gloves.
   • Open options are research-grade and often legally restricted: HOMIE is gated and non-commercial, AirExo-2 is CC BY-NC-SA, GeoRT is CC BY-NC, Aero's CAD is non-commercial, Yale OpenHand is CC BY-NC, and HOPEJr has no licence.
   • DexUMI needs a new exoskeleton for each target hand.
   • Vision capture (Vision Pro, Quest, MediaPipe) is supported by LEAP v2, ORCA and Aero, but records no force.
   • Opening: a commercially licensable glove or exoskeleton that records both force/haptics and kinematics and retargets to many hands.
6. There is a data-scale gap between grippers and hands.
   • Gripper capture has reached 100k trajectories (FastUMI-100K, about 600 h).
   • Open dexterous projects report about 40–100 demos per task (DOGlove, RUKA-v2), and few include synchronised touch (RAPID does).
7. Hands differ widely in shape and control.
   • Actuator count ranges from 7 (Aero) to 22 (Sharpa); some hands have 4 fingers and some 5; joint conventions differ.
   • Retargeting has to be redone for every hand, and there is no standard hand-data schema. LeRobot is the closest thing to a common data and teleop layer.
8. The price gap is closing.
   • Mid-price closed hands now cost about the same as assembled open hands: AgiBot OmniHand $3,775 and Inspire about $5–6.5k, versus ORCA at $3.5–7.1k assembled.
   • So the case for open hands is now modifiability and licence, not price.

8. GitHub repos referenced (stars via GitHits metadata, 2026-10-02)