Source-linked AI summary
LEAP Hand: Low-Cost, Efficient, and Anthropomorphic Hand for Robot Learning
Kenneth Shaw, Ananye Agarwal, Deepak Pathak
TL;DR
Dexterous manipulation has produced limited real-world results, partly because suitable hardware is scarce. The paper introduces LEAP Hand, a low-cost anthropomorphic hand with a kinematic mechanism that preserves dexterity across finger poses, and demonstrates its use in real-world robot learning tasks. LEAP Hand supports teleoperation, behavior cloning, and sim2real, including blind in-hand cube rotation, while offering a low-cost, durable research platform.
Problem
Real-world dexterous manipulation remains limited compared with simulation, motivating suitable hardware for robot learning.
Method
The paper introduces LEAP Hand, combining anthropomorphic design with a universal abduction-adduction mechanism that retains all degrees of freedom across finger positions.
Results
LEAP Hand is used for real-world teleoperation, behavior cloning, and sim2real tasks, including blind in-hand cube rotation, and can continuously hold a grasp for an hour with only a small angle error.
Takeaways & Limitations
The released hardware, simulators, models, instructions, and APIs provide a low-cost development platform for dexterous robot-learning research.
Abstract
from arXiv · showhide
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/
I. INTRODUCTION
LEAP Hand addresses the limited real-world progress in dexterous manipulation by providing a low-cost, robust, anthropomorphic platform for robot learning. Its new kinematic mechanism preserves finger dexterity across poses and supports real-world learning tasks.
- Real-world dexterous manipulation has lagged behind simulation and has mostly been demonstrated with simple one-degree-of-freedom parallel-jaw grippers.
- Hardware cost, maintenance, and limited kinematic structures have restricted access to hands capable of complex dexterous tasks.Shadow costs over 100K USD, while Allegro is described as unreliable, difficult to repair, and non-anthropomorphic.
- LEAP Hand uses off-the-shelf or 3D-printed parts, assembles in under 4 hours for 2000 USD, and is 1/8th the cost of Allegro.It is also designed for in-house repair with a standard $250 3D printer.
- Its design aims to combine anthropomorphism with dexterity so human demonstrations and human-designed environments can support learning.
- The proposed mechanism retains all degrees of freedom across finger positions and is intended to improve grasping and in-hand manipulation.The paper also integrates the hand with video-based teleoperation, behavior cloning, and sim2real, including blind in-hand cube rotation.
II. RELATED WORK
Prior dexterous hands demonstrate capable manipulation but face cost, maintenance, calibration, simulation, or actuation limitations. Related learning work connects robot hands with human-hand data, video teleoperation, and simulation-based policy training.
- Shadow and ADROIT enabled complex contact-rich dexterous tasks but cost about 100K USD and require constant maintenance.
- Inmoov is 3D printable and human-like but has one degree of freedom per finger and tendon actuation that is difficult to calibrate.
- Rapid Manufacturing: Additive manufacturing offers rapid prototyping and is used for LEAP Hand parts including the hard palm and soft rubber fingertips.
- Learning Dexterity: Earlier dexterous manipulation studies include in-hand rotation, valve repositioning, real-world Baoding Ball rotation, and pipe insertion.
- Robot Learning: Related learning methods use human-hand parameters, real-time video teleoperation, web video poses, and large-scale video pre-training.
III. KINEMATIC DESIGN AND ANALYSIS
LEAP Hand’s kinematic design targets human-like structure and dexterity for learning from human data and performing in-hand manipulation. Its universal mechanism preserves abduction-adduction across finger poses, addressing limitations in prior direct-driven hands.
- A hand’s kinematic structure is its joint arrangement, which determines available poses and forces of motion.
- Human fingers combine a two-degree-of-freedom MCP ball joint with one-degree-of-freedom PIP and DIP hinge joints, while the opposable thumb supports power and precision grasps.
- Tendon-driven hands can reproduce flexible ball-joint designs but are expensive and difficult to maintain, whereas direct-driven hands are cheaper but kinematically limited.
- Allegro and LEAP-C Hand lose one degree of freedom in either extended or closed positions because their abduction-adduction motor axis is fixed to the palm.
- B. Evaluating Manipulability via Thumb Opposability: Thumb opposability is evaluated through intersecting thumb-and-finger workspaces, with LEAP Hand reported to have an even spread and a very large contact area.
- A. Universal Abduction-Adduction Mechanism: LEAP Hand moves the abduction-adduction axis into the first finger joint’s frame and keeps it perpendicular, retaining all degrees of freedom at all MCP positions.This provides abduction-adduction in extension and pronation-supination in flexion.
B. Evaluating Manipulability via Thumb Opposability
LEAP Hand is evaluated for dexterity through thumb opposability and manipulability across finger configurations. Its design preserves fingertip mobility across poses and yields practical grasping advantages.
- Thumb opposability: LEAP Hand is compared with Allegro and LEAP-C Hand using thumb–finger workspace intersection and a thumb opposability metric.The comparison samples joint configurations where the thumb and fingers touch to quantify the reachable contact volume.
- Manipulability: LEAP Hand has larger manipulability ellipsoids and greater volumes for both Cartesian and angular motion in three key poses.The evaluated poses include down, fully up, and halfway or curled configurations.
- Kinematic comparison: LEAP Hand retains large fingertip motion in both flexed and extended positions, unlike LEAP-C Hand and Allegro.LEAP-C Hand has greater motion when extended, while Allegro has greater motion when flexed.
- Practical benefits: The increased dexterity supports tighter grasping and faster blind in-hand cube rotation than Allegro.These practical benefits are reported in the grasping and contact-rich cube-rotation tasks.
IV. HAND DESIGN PRINCIPLES
LEAP Hand’s hardware design prioritizes low cost, repairability, robustness, and anthropomorphic dexterity. It uses accessible manufacturing methods and modular construction to support research use.
- Design goals: A suitable dexterous hand should be low-cost, easy to repair, robust, durable, repeatable, versatile, and ideally anthropomorphic.These properties are motivated by the limited accessibility and reproducibility of existing dexterous-manipulation hardware.
- Design goals: Existing ShadowHand and AllegroHand systems cost 100K and 16K USD, respectively, and require manufacturer repair after damage.Their cost and maintenance requirements restrict access for many researchers.
- Modularity: Its modular design permits changes to finger count, finger length, and palm spacing while simplifying repair with few distinct parts.These changes can be made for particular learning tasks or analysis.
B. Robustness
LEAP Hand is designed to withstand demanding robot-learning workloads while producing substantial torque. The paper evaluates material choices, motor capability, and long-duration load holding.
- Robustness requirements: Robot-learning workloads can repeatedly expose a hand to collisions, while heavy-object and tool use require large torques.The hand is therefore expected to keep functioning reliably under harsh treatment.
- Hardware choices: LEAP Hand uses reinforced off-the-shelf plastic brackets instead of costly custom machined metal parts.Only the palm and smaller wire-guide spacers are 3D printed.
- Hardware choices: Its motors are geared for high torque while maintaining hand-like joint speeds of around 8 rad/sec.The design maximizes motor mass within a human-like form factor and supports current or torque limiting.
- Endurance: LEAP Hand continuously holds a 2kg weight on one fingertip for one hour with only a small angle error.Current usage stabilizes with motor temperature, and the top motor reaches 250mA, less than half the maximum possible current.
2) Repeatability test
The repeatability test compares LEAP Hand and Allegro during continuous grasping over one hour. LEAP Hand maintains low joint error, whereas Allegro begins failing after 15 minutes.
- Test setup: The experiment repeatedly raises and lowers a 25g plush dice at 5Hz for one hour by commanding a single base finger joint.The measured quantity is the difference between desired and actual joint angle over time.
- Evaluation: The endurance comparison evaluates both consistency and accuracy by graphing desired-versus-actual joint-angle error over time.This connects repeatability performance to sustained grasping operation.
- LEAP Hand: LEAP Hand maintains consistent errors of 0.025 radians upward and 0.005 radians downward.The paper relates these errors to the PID controller and 750mA current limit.
- Allegro comparison: Allegro begins failing after 15 minutes and completely fails to move on one out of three grasps.The reported cause is motor overheating from continuous grasping strain, not position-sensor failure.
- Related robustness measure: The pullout test separately assesses resistance to outward force, with failure defined by slipping or more than 15 degrees of deviation.Its returned force is used as a grip-strength correlate rather than as a repeatability measure.
V. FABRICATION AND SOFTWARE
LEAP Hand combines rapid fabrication, versatile control modes, simulation support, and evaluations spanning grasping, teleoperation, and learning tasks.
- Fabrication: The hand is assembled from 3D-printed parts, motors, brackets, and cabling in around 4 hours.Fabrication uses a consumer-grade FDM printer and PLA plastic.
- Software: Four control modes support position, current, current-based position, and velocity commands.Current-based position control caps maximum current and torque while following position commands.
- Simulation: The released Isaac Gym and PyBullet simulators support designing and evaluating LEAP Hand versions and sim2real transfer.The simulator’s faithfulness is verified through sim2real experiments.
- Evaluation: The evaluation includes grasping, teleoperation, behavior cloning, and sim2real in-hand manipulation.The experiments compare LEAP Hand and Allegro across machine-learning tasks after an initial grasping test.
- Evaluation: LEAP Hand outperforms or matches Allegro on 9/10 teleoperation tasks.Table IV reports success rate and average completion time for trained operators.
- Evaluation: LEAP Hand’s morphology and strong motors enable cigarette and flat-hand cupping grasps under perturbation testing.The grasping evaluation measures resistance to perturbation force in newtons.
A. Grasping Test using Teleoperation
Teleoperation uses human hand input to control LEAP Hand and collect demonstrations, while its human-like morphology supports direct joint-angle mapping.
- Grasping Test using Teleoperation: Human teleoperation uses VR-glove feedback to explore grasp poses and measure perturbation forces up to 20N.The test covers everyday objects and different power and precision grasps.
- Grasping Test using Teleoperation: LEAP Hand grasps all tested objects and performs many power and precision grasps.Allegro lacks some grasps because of weaker motors and missing adduction/abduction in the extended position.
- Teleoperation: Teleoperation retargeting minimizes an energy function measuring distance between human hand poses and scaled robot hand poses.The formulation uses manually defined palm-to-fingertip vectors, and an MLP is trained to implicitly minimize the energy.
- Teleoperation: LEAP Hand outperforms Allegro on 9/10 teleoperated tasks through direct human-to-robot joint mapping.The paper attributes easier control to LEAP Hand’s morphology, accuracy, and responsiveness.
C. Behavior Cloning from Demonstrations
Behavior cloning combines internet human-video pretraining with a small number of teleoperated demonstrations, and LEAP Hand outperforms Allegro across most evaluated pairs.
- C. Behavior Cloning from Demonstrations: VideoDex and NDP pretrain policies on Epic-Kitchens internet videos before fine-tuning with demonstrations.The demonstrations include prior Allegro Hand data mapped to LEAP Hand.
D. Sim2Real In-Hand Manipulation
The sim2real experiment trains a recurrent policy in Isaac Gym for vision-free cube rotation using proprioceptive history, then evaluates the learned behavior on the real hand.
- D. Sim2Real In-Hand Manipulation: The policy rotates a cube about an axis perpendicular to the palm while inferring cube pose from joint-angle history alone.The contact-rich task does not directly expose the cube pose.
- D. Sim2Real In-Hand Manipulation: A GRU policy is trained with PPO and BPPT in Isaac Gym using rotation rewards and penalties for grasp deviation, work, torque, and object motion.The rotation reward clips vertical angular velocity, while the penalties regularize stability and effort.
- D. Sim2Real In-Hand Manipulation: LEAP Hand achieves faster simulated cube rotations than Allegro.Its joint structure supports the cube from the sides, whereas Allegro periodically releases it to reorient.
- D. Sim2Real In-Hand Manipulation: The paper demonstrates sim2real transfer alongside teleoperation and behavior cloning, and releases the hand’s development platform.The conclusion also identifies low-cost touch sensors as future work.