# Dexterous Humanoid Hands: Landscape, Benchmarks & How to Reproduce SOTA

Oct 2, 2026 · @julian gilyadov

The status quo to beat is an open, low-cost hand (ORCA or LEAP) trained by glove teleop, GPU sim-to-real RL, or pretraining on human video. The best real results now come from human-data pretraining, while real in-hand reorientation on cheap hands still manages only a few consecutive goals.

To know whether our work improves on this, reproduce four anchors on one hand in about eight weeks: simulated reorientation, real reorientation, glove-teleop imitation learning, and POMDAR plus a human-data baseline. Freeze them in the baseline sheet, then test our data, simulator or environment on the metric that matches it.

The biggest openings: nobody has published autonomous POMDAR scores or ORCA reorientation numbers, and real reorientation on low-cost hands lags simulation badly.

## Hardware landscape

Labs train on cheap open hands (LEAP, RUKA, ORCA: about $200–2,000 in parts), while 20–22 DoF tactile hands (Sharpa, Wuji, Tesla) set the 2026 frontier but publish few comparable benchmarks.

**Research hands you can buy or build** (sorted by price)

| Hand | Released | Actuated DoF | Actuation | Tactile | Price (USD) | Known for |
| --- | --- | --- | --- | --- | --- | --- |
| [LEAP Hand v2](https://roboticsproceedings.org/rss21/p132.html) (CMU, 2025) | Jun 2025 (RSS paper) | 8 | Hybrid rigid-soft, servos | None reported | \~200 parts | Under 2 h assembly |
| [RUKA](https://arxiv.org/pdf/2504.13165) (NYU, 2025) | Apr 2025 (arXiv) | 11 (15 joints) | Tendon, forearm motors | None | \~1,300 parts | Learned controllers from MANUS-glove data; Kapandji 10/10; 29 of 33 GRASP-taxonomy grasps |
| [LEAP Hand v1](https://www.roboticscenter.ai/learn/dexterous-robot-hand-comparison-2026) (CMU, 2023) | Jul 2023 (RSS paper) | 16 (4 fingers) | Servos in joints | None | \~2,000 parts; 2,066 kit | Most-used low-cost research hand |
| [ORCA](https://arxiv.org/html/2504.04259v2) (ETH Zurich, 2025) | v1: Apr 2025 (arXiv); v2 line (lite, standard, Touch): Mar 15, 2026; v2.1: no public date found | 17 incl. wrist | Tendon, forearm motors | Fingertip FSR; Touch: 351 taxels | \~2,000 parts; 3,500+ assembled | Published reliability data, LeRobot stack |
| [Wuji Hand](https://www.roboticscenter.ai/learn/dexterous-robot-hand-comparison-2026) | Sep 18, 2025 (Hand 2 teased Jun 2026) | 20 | Micro-actuator per joint | None standard | 16,000 US list | 580 g, 15 N fingertip, 300k+ grasp-cycle rating |
| [Inspire RH56DFX / BFX](https://www.roboticscenter.ai/learn/dexterous-robot-hand-comparison-2026) | 2022 (DFX); BFX date not published | 6 active (12 joints) | Linear servos | Optional fingertip force | 20,599–21,599 US list | Common on Unitree humanoids |
| [Allegro Hand](https://www.roboticscenter.ai/learn/dexterous-robot-hand-comparison-2026) | 2012; V4 2018, V5 2024 | 16 (4 fingers) | Motors in joints, CAN bus | None standard | 21,450 US list | Most-cited hand; HORA and DexPoint baselines |
| [Sharpa Wave](https://mikekalil.com/blog/sharpa-north-ces-2026/) | Oct 2025 (first shipments) | 22 | Proprietary | 1,000+ taxels per fingertip at 180 Hz | Not disclosed | 30 N fingertip force; mass production claimed |
| [Shadow Dexterous Hand](https://en.wikipedia.org/wiki/Humanoid_hand) | 2005 | 20 (24 joints) | Tendon | Extensive | > 100,000 CHF | OpenAI Dactyl and Rubik's cube |

US list prices for LEAP v1, Wuji, Inspire and Allegro are one reseller's (Robotics Center, July 2026); Shadow's cost is from the [ORCA paper](https://arxiv.org/html/2504.04259v2).

Release dates: research hands use the first paper date; commercial hands use first sale or first shipment ([Wuji](https://aiwiki.ai/wiki/wuji_hand), [Inspire](https://www.originofbots.com/humanoid-robot-hand/rh56dfx-by-inspire-robotics-robot-hand-specifications), [Allegro](https://aiwiki.ai/wiki/allegro_hand), [Sharpa](https://interestingengineering.com/ai-robotics/sharpas-advanced-robotic-hand-enters-mass-production), [Shadow](https://aiwiki.ai/wiki/shadow_robot), [ORCA v2 line](https://robohorizon.com/en-us/news/2026/03/orca-dexterity-drops-3-open-source-robotic-hands-starting-at-1500/)). ORCA v2.1 appears in no public release note or the orca\_core repository, so its date is open.

**Humanoid OEM hands** (not sold separately; reported specs as of April 2026, [Wikipedia](https://en.wikipedia.org/wiki/Humanoid_hand))

| Robot hand | Released | DoF | Actuation | Sensing |
| --- | --- | --- | --- | --- |
| Tesla Optimus Gen 3 | 22-DoF hand shown Oct 2024; full Gen 3 robot not revealed as of Apr 2026 | 22 | Tendon; 25 actuators in the forearm | Vision-first; tactile little disclosed |
| Sanctuary Phoenix | 21-DoF hand Dec 2024; Phoenix Gen 7 Apr 2024 | 20–21 | Hydraulic | Micro-barometer pads, 5 mN sensitivity |
| Figure 02 (Helix 02) | Aug 6, 2024 | 16 | Electric unit per finger | Fingertip tactile (\~3 g), palm cameras |
| Fourier GR-2 | Sep 30, 2024 | 12 | Electric | 6 tactile arrays per hand |
| Unitree H2 (Dex5 option) | Oct 20, 2025 | 10–12 | Electric | Optional tactile |
| Apptronik Apollo (PSYONIC Ability) | Apollo Aug 2023; Ability Hand 2021 | Low; prosthesis-derived | Electric | Multi-touch |

DoF count is a weak proxy for dexterity. A May 2026 kinematic metric across 16 hands ([KaRMA](https://arxiv.org/pdf/2605.15548)) ranked LEAP highest and Ability lowest, and found DoF alone does not predict reachable workspace.

Release dates are first public unveilings ([Optimus](<https://en.wikipedia.org/wiki/Optimus_(robot)>), [Sanctuary hand](https://techcouver.com/2024/12/12/tech-milestone-vancouver-robotics-enhanced-dexterity/), [Phoenix Gen 7](https://techcouver.com/2024/04/25/sanctuary-ai-generation-phoenix-robot-carbon-ai/), [Figure 02](https://www.therobotreport.com/figure-02-humanoid-robot-is-ready-to-get-to-work/), [GR-2](https://www.adnkronos.com/immediapress/eng/fourier-unveils-the-next-generation-humanoid-robot-gr-2_7amoXuJIej0JlSeOx5XoCi), [H2](https://technode.com/2025/10/20/unitree-unveils-h2-humanoid-robot-with-lifelike-design/), [Apollo](https://www.therobotreport.com/apptronik-unveils-apollo-humanoid-robot/), [Ability Hand](https://mikekalil.com/blog/apollo-ability-hand/)). Benchmark and dataset dates elsewhere are first paper or release dates ([Elliott & Connolly](https://dblp1.uni-trier.de/pid/234/1097.html), [Isaac Lab](https://www.aiwiki.ai/wiki/isaac_lab), [EgoDex](https://arxiv.org/abs/2505.11709), [ActionNet](https://www.yicaiglobal.com/news/chinas-fourier-makes-humanoid-robot-dataset-open-source), [RoboMind](https://arxiv.org/abs/2412.13877v3), [RealDex](https://arxiv.org/abs/2402.13853v2), [BrainCo Revo2](https://technode.global/?p=107988)).

## ORCA deep dive

ORCA is the best-documented low-cost anthropomorphic research hand: 17 DoF, tendon-driven, under 2,000 CHF in parts, with published reliability, accuracy and learning results ([ORCA paper, IROS 2025](https://arxiv.org/html/2504.04259v2)). Since June 2026 it also has an open learning stack that plugs into Hugging Face's LeRobot ([ORCA platform paper](https://arxiv.org/abs/2606.14561)).

**Design (v1)**

- Kinematics: 16 finger joints plus 1 wrist joint. Fingers 2–5 have MCP, PIP and abduction joints (no DIP); the thumb adds a CMC joint and sits at 15° supination for opposition.
- Actuation: each joint is pulled by a flexor/extensor pair of 0.4 mm braided nylon tendons over metal pins. Motors sit in a forearm "tower" with fans; the wrist is belt-driven with 60° of flexion and extension.
- Reliability features: pin joints that pop out instead of breaking, ratchet spools for re-tensioning in seconds, and auto-calibration that drives each joint to its stops to fit a motor-to-joint ratio without joint encoders.
- Sensing: binary force-sensing resistors under silicone skin on all five fingertips.

**Published data points**

| Metric | Result | Conditions |
| --- | --- | --- |
| Build | < 2,000 CHF materials, < 8 h assembly by one person | DIY, 3D-printed PLA |
| Continuous grasping | 2,250 grasp cycles in 2.5 h, no failure | Grasp every 4 s, wrist cycle every 16 s; stopped by choice |
| Durability claim | > 10,000 cycles (\~20 h) without hardware failure | Abstract figure |
| Joint tracking | Accuracy similar to LEAP Hand; mean latency < 0.2 s | 2 Hz and 5 Hz sine, AprilTags filmed at 60 fps; latency mostly software |
| Sim-to-real RL | Zero-shot tennis-ball reorientation about a given axis | IsaacGymEnvs, 4,096 parallel envs, A2C, \~1 h training with domain randomization |
| Imitation learning | 214 demos (\~2.5 h) gave a policy that ran 7 h 17 min (\~2,000 grasps) with no hardware intervention | Rokoko-glove teleop, Franka arm, 3 cameras, diffusion transformer, 500 epochs ≈ 4 h on one RTX 4090 |
| IL ablation | Masked-cube input beat raw RGB | 60 trials, 10 per table sub-area |
| Teleop skills | Cube stacking, jar-cap twist, fidget-spinner spin, writing "Hello", pouring 50 ml | Qualitative, gloves |
| Tactile limits | Detects 0.05 N; skin wear after \~2,000–4,000 cycles; sensor wires snapped after \~4,500–7,000 | v1 binary FSR design |

**Buying options (Jul–Oct 2026)**

| Option | Price (USD) | Contents |
| --- | --- | --- |
| [BOM kit](https://shop.orcahand.com/products/assembly-material) | 643 | Tendons, connectors; no motors, prints or silicone |
| [Assembly kit with motors](https://shop.orcahand.com/products/orca-assembly-kit) | 4,234 per hand | Dynamixel motors, tested prints, cast skin, fans |
| [Fully assembled v1](https://shop.orcahand.com/products/fully-assembled-orca-hand) (2025) | 5,929 per hand | Tested, hard case, spares |
| [Assembled ORCA Hand, current listing](https://robozaps.com/products/orca-hand) | from 3,500 (Feetech) or 4,500 (Dynamixel) | Manufacturer-confirmed Jul 31, 2026 |
| ORCA Hand Touch (Mar 2026) | 6,100–7,100 | 351 Hall-effect taxels, 6D force per taxel, \~1.25 kg |
| 9-DoF lite (Mar 2026) | BOM < 900 | Coupled tendons, self-build only |

The seller is ORCA Dexterity, Inc., a Delaware company with its team in Zurich; no grip-force or payload figure is published ([RoboZaps record](https://robozaps.com/products/orca-hand)).

**Software**

- [orca\_core](https://github.com/orcahand/orca_core): MIT-licensed Python controller with tension, calibrate and neutral scripts. It auto-detects Dynamixel or Feetech motors and ships v1 and v2 hand configs (\~590 stars, 150 forks).
- Teleop: Rokoko gloves and Apple Vision Pro are supported out of the box; the retargeting code is open ([shop page](https://shop.orcahand.com/products/fully-assembled-orca-hand)).
- [ORCA learning stack](https://huggingface.co/papers/2606.14561) (June 2026): one interface for control, simulation, consumer-device teleop and retargeting. Its reference workflow is VR-headset teleop of in-hand reorientation, a LeRobot-trained policy, and a reproducible evaluation setup.

What ORCA has not published: grip force, payload, YCB-style grasp success rates, or consecutive-success counts for reorientation. Those gaps are exactly where a new data or simulation product can show a measurable gain.

## How dexterous hands get their skills

Five training routes dominate in 2026, and the fastest-moving one is human data: egocentric video or wearable capture, converted into robot trajectories and co-trained with a few robot demos.

| Route | How it works | Typical cost | Main limitation |
| --- | --- | --- | --- |
| Teleop + imitation learning | An operator drives the hand with gloves or a VR headset; a policy (Diffusion Policy, ACT, π0-style VLA) clones the demos, e.g. [ORCA](https://arxiv.org/html/2504.04259v2) (Apr 2025) | 50–200+ robot demos per task | Slow and tied to one hand; tracking quality caps data quality (gloves beat headsets under occlusion) |
| Sim-to-real RL | Train in a GPU simulator with domain randomization, deploy zero-shot, e.g. [MuJoCo Playground](https://arxiv.org/html/2502.08844v1) (Feb 2025) | Minutes to days of GPU time | Needs object-pose tracking and reward design; real results trail simulation |
| Human video to robot | Retarget hand poses from egocentric video into robot trajectories, then pretrain, e.g. [UniDex](https://arxiv.org/html/2603.22264v1) (Mar 2026), [EgoScale](https://arxiv.org/html/2602.16710) (Feb 2026) | Thousands of hours of video | Kinematic and visual gaps; still needs some robot data |
| Human–robot co-training | Wearable rigs (gloves, head cameras) record humans; a few robot demos anchor the embodiment, e.g. [DexWild](https://arxiv.org/html/2505.07813v2) (May 2025) | Human demos collect several times faster than robot demos | Success collapses without any robot demos |
| World models and dexterous VLAs | Pretrain dynamics or policies on mixed human and robot data, then plan or fine-tune per hand, e.g. [DexWM](https://arxiv.org/html/2512.13644v2) (Dec 2025) | Large pretraining, small fine-tune | Cross-hand transfer is still weak |

Two caveats recur across routes: human data alone does not finish the job, and policies rarely transfer across hands without retraining. The State of the art section below gives the numbers.

**Where hands are used today**

- Research benchmarks: in-hand reorientation, grasping, tool use (scissors, spray bottles, kettles) and in-hand assembly.
- Humanoid demos: Sharpa's robot assembled a paper windmill in 30+ autonomous steps at CES 2026 ([report](https://mikekalil.com/blog/sharpa-north-ces-2026/)). OEMs publish videos far more often than success rates.
- Data collection: teleop rigs (ORCA supports Rokoko gloves and Apple Vision Pro) feed imitation-learning datasets.
- Prosthetics-derived hands: PSYONIC Ability (2021) and BrainCo Revo2 (Sep 2025) hands carry prosthetic designs onto humanoids.

## Benchmarks and metrics

There is no single dexterity leaderboard. A credible claim stacks three levels: a hand-level benchmark (POMDAR), a simulation suite (MuJoCo Playground), and a real-world policy protocol with fixed trial counts and reported variance.

**Hand-level benchmarks** (what the hand plus controller can do)

| Benchmark | Measures | Protocol and score | Status |
| --- | --- | --- | --- |
| [POMDAR](https://arxiv.org/html/2604.09294v1) (ETH Zurich, Apr 2026) | 12 in-hand manipulation and 6 grasp tasks from the Elliott & Connolly and GRASP taxonomies | 3D-printed scaffolds; score = 0.8 × correctness + 0.2 × speed vs a human baseline; 20 trials per task; MuJoCo twin | ORCA scored by teleop only (1,140 trajectories, \~25 h); gloves beat Apple Vision Pro on occluded tasks. No autonomous-policy scores yet |
| Elliott & Connolly benchmark (CMU, Jul 2021) | 13 in-hand patterns using the digits only | YCB objects, visual tracking | No grasp tasks |
| HD-marks (2020) | 50 tasks: GRASP grasps, Kapandji thumb postures, in-hand axes | Mostly binary success | Broad, weak cross-lab comparability |
| Kapandji test (1986) + GRASP taxonomy (2016) | Thumb opposition (0–10) and 33 grasp types | Static postures | RUKA: 10/10 and 29 of 33 |
| [KaRMA](https://arxiv.org/pdf/2605.15548) (May 2026) | Kinematic workspace for fine manipulation | Computed from hand models | 16 hands ranked |
| [DexBench](https://www.weforum.org/stories/all/why-hand-dexterity-remains-a-barrier-to-automation/) (RLWRLD, Jun 2026) | 18 atomic industrial tasks in \~80 cases, placed on object-complexity and dexterity-regime axes | Real-world, customer-derived | Endorsed by Lotte, SK Telecom, CJ Logistics and others |

The Elliott & Connolly and HD-marks rows are as summarized in the POMDAR paper.

**Simulation suites**

| Suite | Contents | Headline metric |
| --- | --- | --- |
| [MuJoCo Playground](https://arxiv.org/html/2502.08844v1) (Feb 2025) | LeapCubeReorient and LeapCubeRotateZAxis; open source, pip-installable | Consecutive successes; \~35 min to train on one RTX 4090 |
| Isaac Gym (2021) / Isaac Lab (Jun 2024) | GPU RL used by DeXtreme and by ORCA's ball-rotation policy | Consecutive successes |
| [Bench2Dex](https://arxiv.org/html/2609.15726) (Sep 2026) | 26 bimanual visuo-tactile tasks, 12 hands, \~1.3K teleop demos | Success, with mandatory seeds, episode counts and termination rules |
| [DexVerse](https://arxiv.org/pdf/2607.08751) (Jul 2026) | Modular multi-task, multi-embodiment suite | Not reviewed here |
| [LabDex](https://arxiv.org/pdf/2608.18618) (Aug 2026) | Hierarchical chemistry-lab skills, simulation and real robot | Skill-to-long-horizon scaling |
| [In-hand assembly](https://arxiv.org/pdf/2609.10137) (Sep 2026) | Two-part assembly inside one hand, Isaac Sim | Goal-reaching error by hand morphology |

**Real-world policy metrics worth reporting**

| Metric | Used by | Definition |
| --- | --- | --- |
| Consecutive successes | DeXtreme, MuJoCo Playground | Goals reached before a drop; median and mean over ≥ 10 trials |
| Task progress + final success | UniDex | Mean stage completion and full-task success over 20 trials |
| Unseen-scene success | DexWild | Success in environments absent from training |
| Zero-shot cross-hand transfer | UniDex | Task progress on a hand never trained on |
| Human–robot exchange rate | UniDex | Human demos needed to replace one robot demo (≈ 2:1) |
| Autonomy hours | ORCA | Hours or cycles without hardware intervention |
| Contact fidelity | [TactiDex](https://arxiv.org/pdf/2607.09190) (Jul 2026) | Agreement of contacts and forces with human demos |

**Datasets**

| Dataset | Size | Content |
| --- | --- | --- |
| [EgoDex](https://arxiv.org/html/2512.13644v2) dataset (Apple, May 2025) | 829 h of 1080p video | Egocentric human manipulation with hand-pose annotations |
| [UniDex-Dataset](https://arxiv.org/html/2603.22264v1) (Mar 2026) | 52K trajectories, 9M frames, 8 robot hands | Retargeted from H2O, HOI4D, HOT3D and TACO |
| ActionNet (Fourier, Mar 2025) | 30K trajectories, 2 hands | Dexterous bimanual teleop |
| RoboMind (Dec 2024) | 19K trajectories, 1 hand | Multi-embodiment teleop |
| RealDex (Feb 2024) | 2K trajectories, 2 hands | Human-like grasping |

ActionNet, RoboMind and RealDex sizes are from the UniDex paper's comparison table.

## State of the art

The strongest 2026 results come from human-data pretraining: UniDex-VLA doubled π0's real tool-use success, and EgoScale showed policy quality scales predictably with hours of egocentric video. Real in-hand reorientation on low-cost hands remains weak.

| Date | Work | Hand | Result |
| --- | --- | --- | --- |
| Sep 2026 | [In-hand assembly](https://arxiv.org/pdf/2609.10137) | Sharpa Wave; Wuji, Allegro, XHand in sim | Sharpa and Wuji performed best; Allegro and XHand failed deep insertion |
| Jul 2026 | [UHAS](https://arxiv.org/pdf/2607.03570) | LEAP | Real cube reorientation averaged 2.0 consecutive goals; multi-hand training scored lower than single-hand |
| Jun 2026 | [ORCA learning stack](https://arxiv.org/abs/2606.14561) | ORCA | First open, LeRobot-native loop: VR teleop, policy training, reproducible evaluation (no headline number) |
| Mar 2026 | [UniDex-VLA](https://arxiv.org/html/2603.22264v1) | Inspire, Wuji | 81% task progress and 76% final success vs π0 at 38% and 35%, with 50 demos per task; zero-shot transfer reached 40% on Wuji and 60% on Oymotion |
| Feb 2026 | [EgoScale](https://arxiv.org/html/2602.16710) | 22-DoF hand | Pretrained on 20,854 h of human video with a log-linear scaling law; average success rate up 54% over no pretraining after \~50 h aligned human and 4 h robot data |
| Jan 2026 | [Sharpa North](https://mikekalil.com/blog/sharpa-north-ces-2026/) | Sharpa Wave | Paper windmill assembled in 30+ autonomous steps (demo, no success rate) |
| Dec 2025 | [DexWM](https://arxiv.org/html/2512.13644v2) | Allegro | 10 of 12 real grasps (\~83%) without robot fine-tuning |
| May 2025 | [DexWild](https://arxiv.org/html/2505.07813v2) | Multiple | 68.5% success in unseen environments, \~4× robot-only; 5.8× better cross-embodiment generalization |
| Feb 2025 | [MuJoCo Playground](https://arxiv.org/html/2502.08844v1) | LEAP | Zero-shot cube reorientation: median 3.5, mean 7.1 consecutive goals over 10 trials (best 27) |
| Oct 2022 | [DeXtreme](https://arxiv.org/pdf/2210.13702) | Allegro | Vision-based reorientation: best run 112 consecutive goals; mean 23.1 when capped at 50; 60 h of training |

Real in-hand reorientation varies about 10× by setup: DeXtreme averaged 23 goals on Allegro, while recent low-cost LEAP results average 2–7. Humanoid OEMs (Tesla, Figure, Sanctuary, Sharpa) publish demos rather than success rates, so their capability cannot be benchmarked from outside.

## Reproduction plan

Reproduce four anchors in about eight weeks with one ORCA or LEAP hand, one arm and one RTX 4090-class GPU: simulated reorientation, real reorientation, glove-teleop imitation learning, and a human-data baseline. Freeze the protocol before testing our own work.

> *Diagram: reproduction roadmap · 4 phases, 4 gates (shown in the PDF version).*

Each gate must pass before the next phase starts; the frozen sheet is what our own work is measured against.

**Phase 0: simulation only (week 1, no hardware)**

1. `pip install playground`, then train `LeapCubeReorient` with Brax PPO using the published settings (100M steps, 8,192 envs). Expect \~35 min on one RTX 4090; run 3 seeds ([MuJoCo Playground](https://arxiv.org/html/2502.08844v1)).
2. Log simulated consecutive goals at 0.1 rad tolerance and wall-clock time to a fixed reward.
3. Load ORCA in MuJoCo through the [ORCA learning stack](https://arxiv.org/abs/2606.14561) and run the same task, so later comparisons are hand-matched.
4. Install POMDAR's MuJoCo version ([project page](https://srl-ethz.github.io/POMDAR/)) and teleoperate a few tasks to validate the pipeline.

Gate 0: reward curves and wall-clock match the paper within about 20%.

**Phase 1: real in-hand reorientation (weeks 2–3, no arm needed)**

5. Build the hand. ORCA assembles in under 8 h; then run `scripts/tension.py`, `scripts/calibrate.py` and `scripts/neutral.py` from [orca\_core](https://github.com/orcahand/orca_core). LEAP v1 takes about 3 h.
6. Copy the Playground rig: palm tilted 20° down on an 80/20 frame, one RealSense D415 overhead, a 7 cm cube, policy at 20 Hz, and DeXtreme's cube-pose detector or AprilTags.
7. Deploy zero-shot. Run 10 trials, ending each when the cube drops or stalls for 30 s. Report every trial plus median and mean consecutive goals at 0.4 rad tolerance.

Gate 1: LEAP matches or beats the published LEAP numbers in State of the art.

**Phase 2: teleop and imitation learning (weeks 3–6, arm required)**

8. Mount the hand on a 7-DoF arm (Franka-class; ORCA uses an ISO 9409-1 flange) with two external cameras and one wrist camera.
9. Teleoperate with Rokoko gloves, calibrating retargeting each session. Apple Vision Pro also works but [scores lower](https://arxiv.org/html/2604.09294v1) when fingers are occluded.
10. Collect \~200 demos of ORCA's self-resetting cube pick-and-place (about 2.5 h), plus 50 demos each for two UniDex-style tool tasks, such as kettle pouring and spray-bottle pressing.
11. Train with LeRobot: Diffusion Policy and ACT as baselines, plus a π0-class VLA fine-tune. Budget \~4 h per policy on one RTX 4090.
12. Evaluate pick-and-place over 60 trials across 6 table zones ([ORCA protocol](https://arxiv.org/html/2504.04259v2)), and each tool task over 20 trials with stage-wise progress ([UniDex protocol](https://arxiv.org/html/2603.22264v1)). Add one multi-hour autonomy run.

Gate 2: multi-hour autonomy without hardware intervention, and tool-task results within about 10 points of the published baselines.

**Phase 3: hand-level score and human-data baseline (weeks 6–8)**

13. Print the POMDAR rig. Score teleop first (20 trials per task) to calibrate against ETH's published ORCA plots, then score autonomous policies, which no one has published yet.
14. Build a UniDex-Cap-style capture rig (Apple Vision Pro plus a RealSense L515 on a printed mount), or start from public EgoDex or UniDex-Dataset data.
15. Co-train human and robot demos on one task. Sweep the human-demo count at fixed robot-demo counts to measure the exchange rate.

Gate 3: a frozen baseline sheet in which every published number is reproduced or its gap explained.

**What you need**

- One ORCA (17 DoF, tendon) or LEAP v1 (16 DoF, servos-in-joints) hand; buying both gives a morphology check. Prices are in the sections above.
- A 7-DoF arm with an ISO 9409-1 flange for Phases 2–3; Phase 1 needs only a frame.
- One RTX 4090-class GPU; every baseline above trains on one.
- RealSense D415 or D435 for in-hand tracking; L515 for UniDex-style RGB-D capture.
- Rokoko or MANUS gloves; Apple Vision Pro optional.
- Open question: arm, glove and camera prices are not covered here.

## Evaluation protocol: proving our work beats the status quo

A claim holds only if, on the frozen protocol above, our data, simulator or environment moves a reproduced baseline number at equal or lower cost, with enough trials to separate signal from noise.

**Pick the metric that matches the product**

| If our work is | Baseline to beat | Primary metric | Win condition |
| --- | --- | --- | --- |
| Glove or wearable data | DexWild co-training; UniDex-Cap | Human demos needed per robot demo; unseen-scene success | Fewer human demos per robot demo than UniDex's ratio, or higher unseen-scene success at equal robot demos |
| Egocentric video data | EgoScale scaling curve; EgoDex; UniDex-Dataset | Real task success after pretraining on matched hours; loss-vs-hours slope | Equal success from fewer hours, or a steeper scaling curve |
| Simulator or sim-to-real pipeline | MuJoCo Playground LEAP; DeXtreme | Real consecutive goals; sim-to-real drop; GPU-hours | Higher real median and mean at equal or fewer GPU-hours |
| Training or evaluation environments | POMDAR-sim; Bench2Dex | Rank agreement between sim and real scores across at least 5 policies | Higher rank correlation with real outcomes than existing suites |

**Rules for a fair comparison**

- Pre-register tasks, objects, success definitions and trial counts before any run.
- Change one variable: same hand, arm, cameras, policy architecture and compute; only the data or environment differs.
- Randomize and blind: interleave policies A and B in random order, and keep the person resetting scenes unaware of which is running.
- Size trials for the effect: 20 trials leave about ±22 points of 95% uncertainty at 50% success. Separating 50% from 70% with 80% power takes about 90 trials per arm.
- Report distributions: per-trial values plus median and mean, because consecutive-goal metrics are heavy-tailed.
- Split in-distribution results from held-out objects, scenes and hands.
- Normalize by cost: report success per hour of data collection and per GPU-hour, since human demos collect faster than robot demos.
- Train at least 3 seeds, and publish seeds, episode counts, termination rules and horizons, as Bench2Dex requires.

**Baseline sheet**

| Measure | Published reference | Reproduced | Ours | Δ (95% CI) | Status |
| --- | --- | --- | --- | --- | --- |
| LEAP real reorientation, median / mean consecutive goals | MuJoCo Playground (2025) |  |  |  | Not started |
| ORCA real reorientation, median / mean consecutive goals | None published |  |  |  | Not started |
| ORCA pick-and-place success, 60 trials | ORCA (2025) |  |  |  | Not started |
| Autonomy hours without hardware intervention | ORCA (2025) |  |  |  | Not started |
| Tool-task progress and final success, 20 trials per task | UniDex (2026) |  |  |  | Not started |
| POMDAR score, teleop | POMDAR (2026) |  |  |  | Not started |
| POMDAR score, autonomous policy | None published |  |  |  | Not started |
| Human demos per robot demo | UniDex (2026) |  |  |  | Not started |
| GPU-hours to train reorientation | MuJoCo Playground (2025) |  |  |  | Not started |

The two "None published" rows are open territory: the first credible autonomous POMDAR and ORCA reorientation numbers would themselves define the status quo.

## Sources

Opened in full:

- [ORCA paper (arXiv 2504.04259v2)](https://arxiv.org/html/2504.04259v2)
- [ORCA platform paper (arXiv 2606.14561)](https://arxiv.org/abs/2606.14561) and its [Hugging Face page](https://huggingface.co/papers/2606.14561)
- [orca\_core repository](https://github.com/orcahand/orca_core)
- [RoboZaps ORCA Hand record](https://robozaps.com/products/orca-hand)
- [Robotics Center dexterous hand comparison, July 2026](https://www.roboticscenter.ai/learn/dexterous-robot-hand-comparison-2026)
- [Wikipedia: Humanoid hand](https://en.wikipedia.org/wiki/Humanoid_hand)
- [Sharpa North at CES 2026](https://mikekalil.com/blog/sharpa-north-ces-2026/)
- [UniDex (arXiv 2603.22264)](https://arxiv.org/html/2603.22264v1)
- [MuJoCo Playground (arXiv 2502.08844)](https://arxiv.org/html/2502.08844v1)
- [POMDAR (arXiv 2604.09294)](https://arxiv.org/html/2604.09294v1)
- [WEF on DexBench, June 2026](https://www.weforum.org/stories/all/why-hand-dexterity-remains-a-barrier-to-automation/)

Cited from search excerpts of the source pages:

- ORCA shop: [fully assembled](https://shop.orcahand.com/products/fully-assembled-orca-hand), [assembly kit](https://shop.orcahand.com/products/orca-assembly-kit), [BOM kit](https://shop.orcahand.com/products/assembly-material)
- [LEAP Hand v2 (RSS 2025)](https://roboticsproceedings.org/rss21/p132.html), [RUKA (arXiv 2504.13165)](https://arxiv.org/pdf/2504.13165), [KaRMA (arXiv 2605.15548)](https://arxiv.org/pdf/2605.15548)
- [DexWild (arXiv 2505.07813)](https://arxiv.org/html/2505.07813v2), [DexWM (arXiv 2512.13644)](https://arxiv.org/html/2512.13644v2), [EgoScale (arXiv 2602.16710)](https://arxiv.org/html/2602.16710), [DeXtreme (arXiv 2210.13702)](https://arxiv.org/pdf/2210.13702)
- [UHAS (arXiv 2607.03570)](https://arxiv.org/pdf/2607.03570), [In-hand assembly (arXiv 2609.10137)](https://arxiv.org/pdf/2609.10137)
- [Bench2Dex (arXiv 2609.15726)](https://arxiv.org/html/2609.15726), [DexVerse (arXiv 2607.08751)](https://arxiv.org/pdf/2607.08751), [LabDex (arXiv 2608.18618)](https://arxiv.org/pdf/2608.18618), [TactiDex (arXiv 2607.09190)](https://arxiv.org/pdf/2607.09190)