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)

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.

Release dates: research hands use the first paper date; commercial hands use first sale or first shipment (Wuji, Inspire, Allegro, Sharpa, Shadow, ORCA v2 line). 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)

DoF count is a weak proxy for dexterity. A May 2026 kinematic metric across 16 hands (KaRMA) ranked LEAP highest and Ability lowest, and found DoF alone does not predict reachable workspace.

Release dates are first public unveilings (Optimus, Sanctuary hand, Phoenix Gen 7, Figure 02, GR-2, H2, Apollo, Ability Hand). Benchmark and dataset dates elsewhere are first paper or release dates (Elliott & Connolly, Isaac Lab, EgoDex, ActionNet, RoboMind, RealDex, BrainCo Revo2).

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). Since June 2026 it also has an open learning stack that plugs into Hugging Face's LeRobot (ORCA platform paper).

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

Buying options (Jul–Oct 2026)

The seller is ORCA Dexterity, Inc., a Delaware company with its team in Zurich; no grip-force or payload figure is published (RoboZaps record).

Software

• 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).
• ORCA learning stack (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.

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). 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)

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

Simulation suites

Real-world policy metrics worth reporting

Datasets

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.

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).
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 and run the same task, so later comparisons are hand-matched.
4. Install POMDAR's MuJoCo version (project page) 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. 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 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), and each tool task over 20 trials with stage-wise progress (UniDex protocol). 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

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

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)
• ORCA platform paper (arXiv 2606.14561) and its Hugging Face page
• orca_core repository
• RoboZaps ORCA Hand record
• Robotics Center dexterous hand comparison, July 2026
• Wikipedia: Humanoid hand
• Sharpa North at CES 2026
• UniDex (arXiv 2603.22264)
• MuJoCo Playground (arXiv 2502.08844)
• POMDAR (arXiv 2604.09294)
• WEF on DexBench, June 2026

Cited from search excerpts of the source pages:

• ORCA shop: fully assembled, assembly kit, BOM kit
• LEAP Hand v2 (RSS 2025), RUKA (arXiv 2504.13165), KaRMA (arXiv 2605.15548)
• DexWild (arXiv 2505.07813), DexWM (arXiv 2512.13644), EgoScale (arXiv 2602.16710), DeXtreme (arXiv 2210.13702)
• UHAS (arXiv 2607.03570), In-hand assembly (arXiv 2609.10137)
• Bench2Dex (arXiv 2609.15726), DexVerse (arXiv 2607.08751), LabDex (arXiv 2608.18618), TactiDex (arXiv 2607.09190)