LEAP Hand: Low-Cost, Efficient, and Anthropomorphic Hand for Robot Learning
TL;DR A four-finger anthropomorphic robot hand built from Dynamixel motors, off-the-shelf brackets and 3D-printed parts in about 4 hours for 2000 USD, 1/8th the cost of Allegro. A new knuckle joint keeps sideways finger motion in every pose, and the hand matches or beats Allegro on 9 of 10 teleoperated tasks.
What they do
Real-world dexterous manipulation is held back by hardware. Shadow Hand costs over 100K USD and needs constant maintenance; Allegro (16K USD) has weak motors, closed-source parts that are hard to repair, and non-human kinematics. LEAP Hand is a four-finger hand built from Dynamixel motors, extruded plastic Robotis brackets and a 3D-printed palm and fingertips. It costs 2000 USD, takes about 4 hours to assemble and weighs 595 g. The authors release the URDF, an Isaac Gym and PyBullet simulator, and a ROS/Python API.
How it works
- Universal abduction-adduction. Direct-driven hands approximate the knuckle (MCP) ball joint with two motors. Allegro and LEAP-C Hand (a baseline in the style of C-Hand) fix the sideways-motion axis to the palm, so it stops working when the finger is extended (Allegro) or flexed (LEAP-C). LEAP Hand mounts that axis on the first finger joint, always perpendicular to it, keeping all degrees of freedom in every pose.
- Human-like joints. Joints are analogous to a human hand’s, so human hand poses map directly to joint angles in teleoperation, without a retargeting energy function.
- Robust, repairable build. Load-bearing parts are off-the-shelf brackets rather than prints; fingers are modular; current-based position control caps torque.
Results
- Thumb opposability volume for the index finger is 1,125,556 mm³ vs 409,135 mm³ for Allegro. Allegro’s angular manipulability volume is zero in all three tested poses.
- Holds 2 kg on one fingertip for an hour, which Allegro cannot. Pull-out force is 19.5 N vs 8.5 N for Allegro.
- Grasps all 13 test objects; matches or beats Allegro on 9/10 teleoperated tasks and on 12/14 VideoDex behavior-cloning train/test pairs.
- Blind cube rotation in simulation: 0.2288 rad/s vs 0.0828 for Allegro; the policy transfers to the real hand.
Limitations
No touch sensing yet (left to future work), and the authors note 2000 USD is still not affordable for all. Pull-out force is below the adult human hand in their own table (26.5 N), and LEAP-C Hand scores slightly higher (21.5 N). Teleoperation results come from one trained operator, and the rotation-speed comparison is in simulation only.
Figures
Abstract
Dexterous manipulation has been a long-standing challenge in robotics. While machine learning techniques have shown some promise, results have largely been currently limited to simulation. This can be mostly attributed to the lack of suitable hardware. In this paper, we present LEAP Hand, a low-cost dexterous and anthropomorphic hand for machine learning research. In contrast to previous hands, LEAP Hand has a novel kinematic structure that allows maximal dexterity regardless of finger pose. LEAP Hand is low-cost and can be assembled in 4 hours at a cost of 2000 USD from readily available parts. It is capable of consistently exerting large torques over long durations of time. We show that LEAP Hand can be used to perform several manipulation tasks in the real world -- from visual teleoperation to learning from passive video data and sim2real. LEAP Hand significantly outperforms its closest competitor Allegro Hand in all our experiments while being 1/8th of the cost. We release detailed assembly instructions, the Sim2Real pipeline and a development platform with useful APIs on our website at https://leap-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.