RUKA: Rethinking the Design of Humanoid Hands with Learning

Anya Zorin, Irmak Guzey, Billy Yan, Aadhithya Iyer, Lisa Kondrich, Nikhil X. Bhattasali, Lerrel Pinto

Seminal April 2025

Teleoperated RUKA on a robot arm picks up a sponge and moves it to a cup.

TL;DR RUKA is an open-source, human-sized, tendon-driven hand with 15 degrees of freedom and 11 actuators that costs under $1,300 to build from 3D-printed and off-the-shelf parts. Its controllers are learned from data recorded by a MANUS motion-capture glove fitted onto the robot hand, and it beats LEAP and Allegro on pinch force, payload (6.0 kg) and slip tests.

What they do

Robot hands trade off precision, compactness, strength and cost. RUKA is an open-source tendon-driven hand with the dimensions of an average human hand, built from 3D-printed parts and off-the-shelf components for under $1,300 in about 7 hours. Tendons make joint positions hard to model, so the authors learn the controllers from data instead of adding joint encoders.

How it works

Hardware. Five fingers with 15 degrees of freedom driven by 11 Dynamixel motors in the forearm: three for the thumb, two for each other finger, with the PIP and DIP joints coupled on one tendon. Tendons are braided fishing line routed through PTFE tubes; springs extend the fingers.
Data collection. A MANUS motion-capture glove is fitted onto the robot hand itself. The hand then runs random walks over its motor positions, recording fingertip positions, joint angles and motor positions at 15 Hz with no human involved.
Controllers. One per finger: an LSTM over the past 10 finger states plus an MLP head predicts motor positions from a desired fingertip position (thumb) or joint angles (other fingers).
New builds. An auto-calibration script finds each motor’s range for the current tendon tension.

Results

Reachability: 10/10 on the Kapandji test (Allegro and LEAP score 9/10) and 29 of the 33 GRASP taxonomy grasps.
Strength: 2.74 N pinch, 6.0 kg payload and 33.02 N DIP/PIP slip force, against 2.45 N, 4.0 kg and 25.17 N for LEAP. It ran 20 hours continuously without a significant drop in motor precision.
Controller: mean thumb fingertip error per axis is 0.20-0.27 cm when replaying robot poses and 0.54-0.83 cm for human poses; MLP and search-based baselines reach about 2 cm on human poses. On a newly built hand, fingertip positions differ by 3 mm on average.
Teleoperation runs at 25 Hz, and HuDOR policies for cube flipping and bread pick-and-drop train in about 45 minutes.

Limitations

Training the controllers needs a MANUS glove, a cost barrier to replication. The hand has no tactile sensing and no abduction at the MCP joints. The plastic-geared motors wear out, and collisions tend to damage the MCP joint; either repair takes under 20 minutes.

Videos

RUKA picks a small cylinder off the table and sets it on a stand.

RUKA mirrors a MANUS glove wearer's hand while holding a marker, a hammer and a disc.

Figures

From Figure 1: RUKA is a tendon-driven humanoid hand whose size and morphology closely match a human hand, enabling p... — From Figure 1: RUKA is a tendon-driven humanoid hand whose size and morphology closely match a human hand, enabling power, precision and fine-grained grasps.

From Figure 3: (A) The joints giving RUKA its 15 degrees of freedom. (B) The splay of the fingers allows natural abdu... — From Figure 3: (A) The joints giving RUKA its 15 degrees of freedom. (B) The splay of the fingers allows natural abduction-adduction movement without an active degree of freedom. (C) The MCP and coupled PIP/DIP tendons are responsible for flexion, while springs are responsible for extension.

From Figure 8: Keypoints from the MANUS gloves (left and right) and the controller architecture (center). An LSTM tak... — From Figure 8: Keypoints from the MANUS gloves (left and right) and the controller architecture (center). An LSTM takes the current and previous 10 fingertip positions, and an MLP head predicts the motor positions for each finger.

From Figure 11: Rollouts of cube flipping and bread pick and drop learned with HuDOR on RUKA. For each task, the bott... — From Figure 11: Rollouts of cube flipping and bread pick and drop learned with HuDOR on RUKA. For each task, the bottom row shows the open-loop rollout of the human trajectory and the top row the final rollout after online finetuning.

Abstract

Dexterous manipulation is a fundamental capability for robotic systems, yet progress has been limited by hardware trade-offs between precision, compactness, strength, and affordability. Existing control methods impose compromises on hand designs and applications. However, learning-based approaches present opportunities to rethink these trade-offs, particularly to address challenges with tendon-driven actuation and low-cost materials. This work presents RUKA, a tendon-driven humanoid hand that is compact, affordable, and capable. Made from 3D-printed parts and off-the-shelf components, RUKA has 5 fingers with 15 underactuated degrees of freedom enabling diverse human-like grasps. Its tendon-driven actuation allows powerful grasping in a compact, human-sized form factor. To address control challenges, we learn joint-to-actuator and fingertip-to-actuator models from motion-capture data collected by the MANUS glove, leveraging the hand's morphological accuracy. Extensive evaluations demonstrate RUKA's superior reachability, durability, and strength compared to other robotic hands. Teleoperation tasks further showcase RUKA's dexterous movements. The open-source design and assembly instructions of RUKA, code, and data are available at https://ruka-hand.github.io/.

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