# ORCA Hand

> Source: https://aiwiki.ai/wiki/orca_hand
> Updated: 2026-07-06
> Categories: Open Source AI, Robot Hardware, Robotics
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/)
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution to "AI Wiki (aiwiki.ai)".

**ORCA** (a backronym for **O**pen-source, **R**eliable, **C**ost-effective, **A**nthropomorphic) is an open-source [dexterous hand](/wiki/dexterous_hand) developed at the Soft Robotics Lab (SRL) of ETH Zurich in Switzerland. It is a tendon-driven, human-scale robotic hand with 17 [degrees of freedom](/wiki/degrees_of_freedom) and integrated tactile sensing that one person can build in under eight hours for a material cost below 2,000 Swiss francs. ORCA was designed to lower the barrier to dexterous manipulation research: the complete design files, control software, simulation environment, bill of materials, and step-by-step assembly and repair guides are released publicly so that any lab can reproduce and repair the hand.[1][2][3] The primary paper, "ORCA: An Open-Source, Reliable, Cost-Effective, Anthropomorphic Robotic Hand for Uninterrupted Dexterous Task Learning," was accepted to the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) 2025.[1][11]

*This article is about the open-source dexterous hand from ETH Zurich. For the unrelated humanoid robot of the same name, see [Cyan Robotics ORCA](/wiki/cyan_robotics_orca).*

## At a glance

| Field | Detail |
| --- | --- |
| Developer | Soft Robotics Lab, ETH Zurich (commercialized via Orca Dexterity) |
| Institution | ETH Zurich, Department of Mechanical and Process Engineering (D-MAVT), Switzerland |
| Lead researcher | Prof. Robert K. Katzschmann |
| Type | Open-source, tendon-driven, anthropomorphic [robotic hand](/wiki/humanoid_robot_hands) |
| Degrees of freedom | 17 (16 in the fingers, 1 in the wrist) |
| Actuation | 17 Dynamixel servo motors in the forearm, antagonistic nylon tendons |
| Sensing | Force-sensing-resistor fingertip sensors (v1); optional high-resolution tactile skins |
| Material cost | Below 2,000 CHF (roughly 2,000 USD) |
| Assembly | One person, under 8 hours |
| License | MIT (software) plus Creative Commons non-commercial (hardware) |
| First release | 2025 |
| Key papers | arXiv:2504.04259 (IROS 2025); arXiv:2606.14561 (2026) |

## Background and origin

ORCA was created by the Soft Robotics Lab at ETH Zurich, part of the Department of Mechanical and Process Engineering and the Institute of Robotics and Intelligent Systems, under principal investigator Robert K. Katzschmann.[2][3] The 2025 paper lists Clemens C. Christoph, Maximilian Eberlein, Filippos Katsimalis, Arturo Roberti, Aristotelis Sympetheros, Michel R. Vogt, Davide Liconti, Chenyu Yang, Barnabas Gavin Cangan, Ronan J. Hinchet, and Katzschmann as authors.[1][11][14]

The motivation was a persistent gap in [robotics](/wiki/robotics) research. Highly capable anthropomorphic hands such as the Shadow Dexterous Hand cost on the order of a hundred thousand dollars and demand specialist maintenance, while cheaper research hands often trade away human-like kinematics, durability, or the tactile feedback needed for manipulation. ORCA targets the middle ground: a hand that is faithful to human hand geometry, cheap enough for a single grant to buy several, and robust enough to run for hours of unattended data collection without a technician resetting it.[1][2] The name encodes those four goals directly: open-source, reliable, cost-effective, and anthropomorphic.[1]

## Design and mechanism

ORCA has 17 active degrees of freedom, of which 16 drive the fingers and thumb and one drives the wrist.[1][3] The four fingers (index through little finger) each carry three actuated joints: a metacarpophalangeal (MCP) flexion joint, a proximal interphalangeal (PIP) joint, and an abduction/adduction (ABD) joint that spreads the finger sideways. The distal joints follow through a coupled linkage rather than a dedicated motor, mirroring how the human distal interphalangeal joint tends to move with the middle joint. The thumb is more richly actuated, with four degrees of freedom (a carpometacarpal joint, an MCP joint, an abduction joint, and an interphalangeal joint), giving it the opposability that human-designed tools assume. Together that is 12 finger plus 4 thumb, or 16 degrees of freedom, and the 17th is the wrist.[3]

Actuation is [tendon-driven](/wiki/tendon_driven). Rather than placing a motor at each knuckle, ORCA houses its servos in the forearm and routes thin tendons across the joints, which keeps the hand itself slim and light and moves the mass proximally, much as human muscles sit in the forearm. Each finger joint is pulled by two antagonistic tendons, one for flexion and one for extension, made from braided nylon fishing line roughly 0.4 mm in diameter; the tendons are routed close to the joint rotation centers to minimize friction.[3] The hand uses 17 Robotis Dynamixel servos: 16 model XC330-T288-T units for the fingers and one larger XC430-T240BB-T for the wrist.[10] The wrist itself is a belt-driven rotational joint offering roughly 60 degrees of flexion and extension.[3] All structural parts are 3D printable, and the hand is sized to approximate an average adult human hand, so it can interact with everyday objects and tools built for people.[1][3]

## Tactile sensing

The first-generation ORCA (v1) integrates contact sensing at the five fingertips using thin-film force-sensing resistors (the RP-C7.6-ST model). These provide effectively binary touch feedback, registering whether a fingertip is in contact and detecting forces the paper reports as low as about 0.05 N, below the sensor's nominal trigger rating.[3] This lightweight scheme keeps cost and wiring simple while still giving learning policies a usable grasp-contact signal, a form of [tactile sensing](/wiki/tactile_sensing) that many low-cost research hands omit entirely.

The 2026 platform release extends the sensing options substantially. Alongside the basic fingertips, ORCA offers higher-resolution tactile skins reported at 51 to 83 taxels per digit, sitting under a soft cast-silicone covering, together with richer joint proprioception. These options let the same hand span everything from simple contact detection to dense tactile imaging for contact-rich manipulation.[6] Because the earlier brief description of ORCA as carrying dense "tactile sensors under a soft silicone skin" conflates the two generations, it is worth noting the distinction: v1 ships with binary fingertip sensors, while the taxel-array skins are an added platform option.[3][6]

## Reliability and maintenance features

Reliability is ORCA's headline design theme, and it is pursued through three ideas.[1][2] First, the joints are built to fail safely: instead of rigid pins that snap under an overload, ORCA uses pop-in pin joints that dislocate when pushed too hard and are simply pressed back into place, so a collision or an over-tensioned tendon costs seconds rather than a replacement part.[2][7] Second, the hand supports auto-calibration, letting it re-establish its joint zero positions in software rather than through manual mechanical adjustment. Third, a manual retensioning ratchet-spool mechanism mounted at the motors lets an operator take up tendon slack quickly, addressing the classic weakness of tendon hands where stretched cables gradually degrade accuracy.[3][8]

Those choices are what let ORCA run for long unattended sessions. The paper documents a continuous-actuation test of roughly 2,250 grasping cycles over about 2.5 hours with no overheating or performance loss, and a separate [imitation learning](/wiki/imitation_learning) session that ran 7 hours and 17 minutes (about 2,000 grasps) without a hardware failure or tendon-slack problem.[3][8] ETH's project page summarizes the durability more broadly, stating the hand withstands more than 10,000 continuous operation cycles, equivalent to roughly 20 hours, without hardware failure.[2] The lab frames this endurance as the point of the "uninterrupted" wording in the paper title: a hand you can leave collecting manipulation data overnight.

## Open-source release and assembly

Everything needed to build an ORCA is published. The release includes the printable STL geometry, a bill of materials with sourcing links, the core Python control library (orca_core), a simulation environment (orca_sim), URDF and MJCF robot descriptions, teleoperation retargeting code, and beginner-oriented documentation, with an illustrated repair guide hosted on iFixit.[3][5][9] The material cost of the parts comes in below 2,000 Swiss francs, and the developers report that a single person can assemble the hand in under eight hours; the structural components are 3D printed and the outer skin is cast in silicone.[1][2] Software is distributed under the permissive MIT license, while the hardware design carries a Creative Commons non-commercial license.[4][5]

For groups that prefer not to source and print parts themselves, the project has a commercial arm, Orca Dexterity, which sells assembly kits (including the motors) and pre-assembled hands at prices above the raw bill-of-materials cost.[4][10] The open files and the paid kits target the same audience of robot-learning labs, with the free route emphasizing reproducibility and repair.

## Teleoperation and dexterous learning

ORCA's anthropomorphic geometry is meant to simplify the mapping from a human hand to the robot, a step known as retargeting, which in turn makes it straightforward to teleoperate the hand and to train on human demonstration data.[1] Out of the box, teleoperation works with Rokoko motion-capture gloves and with an Apple Vision Pro headset, so an operator can puppet the hand directly.[2][3] The 2026 platform paper broadens this to camera-based hand tracking via MediaPipe, the Meta Quest 3 headset, and Manus motion-capture gloves, with real-time retargeting reported at 40 to 50 frames per second.[6]

The hand is designed to support the full modern manipulation-learning pipeline. The team demonstrated collecting teleoperated demonstrations and training policies by behavioral cloning, as well as zero-shot sim-to-real [reinforcement learning](/wiki/reinforcement_learning): a policy trained for about an hour in simulation with domain randomization transferred directly to the physical hand for an in-hand tennis-ball reorientation task.[1][3] The follow-up platform work, "orca: A Platform for Open-Source Dexterity Research" (arXiv:2606.14561, 2026), packages this into a unified software stack (orca_core, orca_sim, orca_arm, orca_teleop) that presents one control abstraction and one data format across the real and simulated hand, integrates full-arm setups such as the Franka Panda, and plugs into the LeRobot ecosystem so that policies like ACT, Diffusion Policy, and pi-zero can be trained with the same tools used for grippers.[6]

## Comparison to other open-source and research hands

ORCA sits in a lineage of anthropomorphic research hands that trade off cost, dexterity, and durability differently.

| Hand | Origin | DOF | Actuation | Approx. cost | Open-source |
| --- | --- | --- | --- | --- | --- |
| ORCA v1 | ETH Zurich SRL | 17 (16 + wrist) | Tendon-driven, forearm servos | < 2,000 CHF (BOM) | Yes |
| LEAP Hand | Carnegie Mellon | 16 | Direct-drive, motors at joints | ~2,000 USD (BOM) | Yes |
| Allegro Hand | Wonik Robotics | 16 | Direct-drive | Tens of thousands USD | No |
| [Shadow Dexterous Hand](/wiki/shadow_robot) | Shadow Robot Company | ~20 to 24 | Tendon (pneumatic or electric) | ~100,000+ USD | No |

The most direct comparison is the LEAP Hand from Carnegie Mellon, another roughly 2,000 USD, open-source, 16-DOF hand that assembles in a few hours.[12][13] The key mechanical difference is that LEAP is direct-drive, placing its Dynamixel motors at the joints, which makes the hand bulkier; ORCA is tendon-driven with the motors in the forearm, giving a slimmer, more human-like hand. In their own benchmarking the ORCA authors reported achieving accuracy similar to the LEAP Hand while being far less bulky.[8] The [Allegro Hand](/wiki/allegro_hand) from Wonik Robotics is the long-standing 16-DOF commercial research standard but costs an order of magnitude more and is not open, and the Shadow Dexterous Hand offers the highest degree count and fidelity at roughly fifty times the price. ORCA also joins a growing set of newer open hardware efforts, including the TetherIA Aero Hand Open from [TetherIA](/wiki/tetheria) and the DexHand from [The Robot Studio](/wiki/the_robot_studio), that share the goal of cheap, reproducible dexterity.

## Reception

ORCA was received as a notable step in democratizing dexterous-manipulation hardware. Robotics and engineering press highlighted the combination of a sub-8-hour build, a roughly 2,000-franc parts cost, and the self-repairing pop-in joints, framing it as a hand a lab can "build in a day" and keep running without a specialist.[7][8] Coverage repeatedly stressed the accessibility angle, that a capable anthropomorphic hand had historically meant either a six-figure Shadow Hand or a compromise, and that ORCA plus the surrounding open software stack lowers that barrier for [robot manipulation](/wiki/robot_manipulation) researchers.[2][7] Acceptance of the paper at IROS 2025, one of the field's major venues, and the subsequent 2026 platform paper co-authored across ETH Zurich, the University of Oxford, and the Orca Dexterity spinoff, indicate uptake within the research community.[6][11]

## Limitations

ORCA's design involves clear trade-offs. The base v1 tactile system is limited to binary fingertip contact sensing; dense tactile imaging requires the optional taxel skins from the later platform release, so out-of-the-box touch information is coarse.[3][6] Tendon-driven actuation, while enabling the slim form factor, introduces cable stretch and friction that the manual ratchet retensioning mitigates but does not eliminate, meaning an operator still periodically retensions the hand.[3][8] The self-dislocating joints prioritize safe failure over rigidity, which can matter for tasks needing high stiffness. The hardware license is non-commercial, so companies cannot freely productize the design, and building the hand, though inexpensive, still requires access to a 3D printer, silicone casting, and several hours of careful assembly. Finally, as a small-team academic platform rather than a certified industrial product, ORCA has not been validated for safety-critical or long-term field deployment.[1][4]

## See also

- [Dexterous hand](/wiki/dexterous_hand)
- [Humanoid robot hands](/wiki/humanoid_robot_hands)
- [Tendon-driven actuation](/wiki/tendon_driven)
- [Tactile sensing](/wiki/tactile_sensing)
- [Shadow Robot Company](/wiki/shadow_robot)
- [Allegro Hand](/wiki/allegro_hand)
- [TetherIA](/wiki/tetheria)
- [The Robot Studio](/wiki/the_robot_studio)

## References

1. ORCA: An Open-Source, Reliable, Cost-Effective, Anthropomorphic Robotic Hand for Uninterrupted Dexterous Task Learning. arXiv:2504.04259. https://arxiv.org/abs/2504.04259
2. The ORCA Hand: Open-Source Dexterity at Your Fingertips. Soft Robotics Lab, ETH Zurich. https://srl.ethz.ch/orcahand.html
3. ORCA paper, full text (HTML). arXiv:2504.04259v1. https://arxiv.org/html/2504.04259v1
4. ORCA Hand project site (Orca Dexterity). https://www.orcahand.com
5. orca_core: Core Python controller of the ORCA Hand (GitHub). https://github.com/orcahand/orca_core
6. orca: A Platform for Open-Source Dexterity Research. arXiv:2606.14561. https://arxiv.org/abs/2606.14561
7. ORCA Hand: ETH Zurich Open-Sources a Dexterous Robotic Hand You Can Build in a Day. Conceptuel Newsroom. https://newsroom.conceptuel.ch/ORCA-Hand-ETH-Zurich-Open-Source-Dexterous-Robotic-Hand/
8. The Next Generation of Robotic Hands. Tech Briefs. https://www.techbriefs.com/component/content/article/54241-the-next-generation-of-robotic-hands
9. ORCA v1 repair guide. iFixit. https://www.ifixit.com/Device/ORCA_v1
10. ORCA v1 Open-Source Humanoid Robotic Hand (specifications listing). Humanoid.guide. https://humanoid.guide/product/orca-v1/
11. ORCA (record and IROS 2025 acceptance). NASA ADS. https://ui.adsabs.harvard.edu/abs/2025arXiv250404259C/abstract
12. LEAP Hand: Low-Cost, Efficient, and Anthropomorphic Hand for Robot Learning. arXiv:2309.06440. https://arxiv.org/abs/2309.06440
13. LEAP Hand project site, Carnegie Mellon University. https://v1.leaphand.com/
14. ORCA (paper record). Semantic Scholar. https://www.semanticscholar.org/paper/611b3e409a34b65a7e557494bc8e316344785abe