ORCA: An Open-Source, Reliable, Cost-Effective, Anthropomorphic Robotic Hand for Uninterrupted Dexterous Task Learning

Clemens C. Christoph, Maximilian Eberlein, Filippos Katsimalis, Arturo Roberti, Aristotelis Sympetheros, Michel R. Vogt, Davide Liconti, Chenyu Yang, Barnabas Gavin Cangan, Ronan J. Hinchet, Robert K. Katzschmann

Seminal April 2025

TL;DR An open-source, 17-DoF tendon-driven anthropomorphic hand with fingertip touch sensors that one person can build in under eight hours for less than 2,000 CHF. It ran an imitation-learning pick-and-place policy for over 7 hours (about 2,000 grasps) with no intervention on the hand hardware.

What they do

Anthropomorphic hands are either expensive and hard to maintain (the Shadow Hand costs over 100,000 CHF) or direct-driven and bulky. ORCA is an open-source, 3D-printable, tendon-driven hand the size of a human hand: 17 DoF including a 1-DoF wrist, about 1.2 kg, under 2,000 CHF in materials, and assembled by one person in less than eight hours. The paper describes the design and tests it for durability, tracking accuracy, teleoperation, imitation learning and sim-to-real RL.

How it works

Tendons. Each finger joint is driven by a flexor and an extensor tendon (0.4 mm braided fishing line) routed over metal pins and through Teflon tubes. Ratchet spools on the motors let the user re-tension a tendon in seconds. The wrist uses a belt drive.
Poppable pin joints. Bearings sit in arc-shaped grooves, so a joint dislocates under excessive load instead of breaking and can be pushed back in.
Auto-calibration. Tendons pass through or near each joint’s center of rotation, making joints roughly decoupled and linear. The hand drives every joint to both mechanical stops and, with the range of motion known from CAD, fits a linear motor-to-joint map. No joint sensors are needed.
Touch. Force-sensing resistors under a silicone skin give binary contact on all five fingertips.

Results

2,250 grasps of a plush toy over 2.5 hours with no breakage, motor shutdown or excessive slack; the test was stopped voluntarily. The hand holds up to 10.5 kg with four fingers and 2 kg with the index finger alone.
Sine tracking at 0.2 and 0.5 Hz is about as accurate as the direct-driven LEAP hand, with average latency below 0.2 s for both.
A diffusion-transformer policy trained on 214 glove-teleoperated demonstrations ran a self-resetting cube pick-and-place for 7 h 17 min (about 2,000 grasps) with no intervention on the hand.
An in-hand tennis-ball reorientation policy, trained for one hour in simulation, transferred zero-shot.

Limitations

Tendons need manual re-tensioning during prolonged use. The tactile sensors were not durable: the silicone skin degraded on two fingertips after roughly 2,000 to 4,000 grasps, and sensor wires snapped on three fingers after 4,500 to 7,000. Touch was not used in the learning experiments, and tracking was not evaluated above 0.5 Hz.

Videos

Figures

From Figure 1: (A) The ORCA hand closely mimics its human counterpart with the same form factor, a bony structure, an... — From Figure 1: (A) The ORCA hand closely mimics its human counterpart with the same form factor, a bony structure, and silicone-cast skin. It is 3D-printed but incorporates joints designed to pop before breaking. (A1) Just before the joint pops. (A2) Applying pressure pops the joint into place and keeps it secure. (A3) The spool system. (B) A self-resetting imitation learning policy running for over 7 hours. (C) The reliability test.

From Figure 3: (A) Naming convention of the joints. The thumb includes an additional degree of freedom. (B) Auto-cali... — From Figure 3: (A) Naming convention of the joints. The thumb includes an additional degree of freedom. (B) Auto-calibration: the three-step process moves all joints to their respective limits and determines a mapping between motor and joint angles without any external sensors. (C) Tendon routing: the tendons are guided around metal pins to reduce friction, and always through the center of rotation.

From Figure 4: (A) Joint response comparison at 0.2 Hz and 0.5 Hz: ORCA achieves accuracy and latency comparable to t... — From Figure 4: (A) Joint response comparison at 0.2 Hz and 0.5 Hz: ORCA achieves accuracy and latency comparable to the LEAP hand despite its tendon-driven design. (B) Accuracy benchmarking setup using AprilTags to infer ground-truth joint angles. (C) Reliability test: 2.5 hours of continuous grasping (2200+ cycles) and wrist motion (550+ cycles), with stable motor current and temperature.

From Figure 6: (A) Prolonged imitation learning experiments over several hours. Walls and a sliding surface allow for... — From Figure 6: (A) Prolonged imitation learning experiments over several hours. Walls and a sliding surface allow for self-resetting of the experiment. (B) A policy trained in simulation deployed zero-shot on the real hand to reorient a tennis ball.

Abstract

General-purpose robots should possess human-like dexterity and agility to perform tasks with the same versatility as us. A human-like form factor further enables the use of vast datasets of human-hand interactions. However, the primary bottleneck in dexterous manipulation lies not only in software but arguably even more in hardware. Robotic hands that approach human capabilities are often prohibitively expensive, bulky, or require enterprise-level maintenance, limiting their accessibility for broader research and practical applications. What if the research community could get started with reliable dexterous hands within a day? We present the open-source ORCA hand, a reliable and anthropomorphic 17-DoF tendon-driven robotic hand with integrated tactile sensors, fully assembled in less than eight hours and built for a material cost below 2,000 CHF. We showcase ORCA's key design features such as popping joints, auto-calibration, and tensioning systems that significantly reduce complexity while increasing reliability, accuracy, and robustness. We benchmark the ORCA hand across a variety of tasks, ranging from teleoperation and imitation learning to zero-shot sim-to-real reinforcement learning. Furthermore, we demonstrate its durability, withstanding more than 10,000 continuous operation cycles - equivalent to approximately 20 hours - without hardware failure, the only constraint being the duration of the experiment itself. Video is here: https://youtu.be/kUbPSYMmOds. Design files, source code, and documentation are available at https://srl.ethz.ch/orcahand.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.

Matt Strong