LEAP Hand V2: Open-Source Robotic Hand

• LEAP Hand V2 is a five-fingered, tendon-driven robotic hand that integrates dense tactile arrays, ToF depth, thermal, and RGB sensors for advanced manipulation.
• It preserves a 12-DoF architecture while enhancing kinematics with modified phalanges, compliance pads, and electromagnetic shielding for improved performance.
• Its three-tier control system—from joint-level MCUs to a central Raspberry Pi running ROS2—enables real-time sensor fusion and closed-loop grasping with high success rates.

The LEAP Hand V2, formally known as the MOTIF Hand, is a five-fingered, tendon-driven robotic hand platform that advances dexterous manipulation capabilities through the integration of multimodal sensing, robust mechanical enhancements, and a three-tier an onboard electronics and control architecture. Extending the original 12-DoF LEAP Hand, the V2 model incorporates onboard dense tactile arrays, a time-of-flight (ToF) depth sensor, a thermal camera, multiple inertial measurement units (IMUs), and an RGB camera while remaining cost-effective (under $4,000 USD) and fully open-source. The MOTIF Hand’s hardware and software enable experimental manipulation tasks that leverage temperature, force, motion, and vision sensing in closed-loop, real-world scenarios (Zhou et al., 24 Jun 2025).

1. Mechanical and Structural Enhancements

The LEAP Hand V2 preserves the underlying architecture of the original LEAP Hand—a 12-DoF, tendon-driven, 5-finger mechanism actuated by Dynamixel servos—but incorporates several significant structural modifications:

• Actuation and Geometry: Dynamixel RX-10 servos and a finger-level four-bar, tendon-driven linkage are retained. The proximal phalanges are lengthened by 5 mm, and the distal phalanges are shortened by 3 mm, accommodating IMU boards at each joint with no net change to reach. The palm is thickened by 2 mm (PLA), permitting integration of a second-tier PCB and the side-mounted ToF sensor.
• Degrees of Freedom: 12 revolute DoFs persist (3 per finger), with a 2-DoF thumb base for abduction/adduction. Compliance pads (0.5 mm silicone) are added around each joint to provide minor passive compliance and reduce electromagnetic interference.
• Materials and Compliance:
  • Load-bearing members: 3D-printed PLA, palm reinforced with a 0.8 mm internal aluminum stiffener.
  • Joints: brass-sleeved revolute bearings.
  • Silicone “valleys” reduce tendon friction and IMU chatter.
  • Palm cavity lined with electromagnetic-shielding tape to minimize interference with BMM350 IMUs.

These design choices preserve the LEAP Hand’s kinematic expressivity while supporting sensor integration and enhanced mechanical robustness (Zhou et al., 24 Jun 2025).

2. Sensor Suite and Multimodal Integration

The MOTIF Hand integrates five sensing modalities:

• Dense Tactile Arrays: Each finger pad carries a 6×6 grid of taxels (2.5 mm²/taxel, 20 g trigger), mounted on both proximal and middle phalanges, sampled at 200 Hz (I²C chain).
• Depth (ToF) Sensor: VL53L5 module, palm-centered, supply a 64×64 depth map ($\pm10$ cm, 15 Hz).
• Thermal Camera: FLIR Lepton 3.5 (raw 160×120, interpolated 1280×960, 8.7–13.5 μm), side-mounted with $\pm15^\circ$ adjustability, outputs $T(u,v)$ at 8 Hz.
• Visual (RGB) Camera: Raspberry Pi Camera Module 2 (8 MP, 30 fps), dorsally mounted for color and segmentation in 3D reconstruction.
• Inertial Measurement Units (IMUs): BMM350 nine-axis modules at every finger joint (11 total) and 4 on the palm. Accel/gyro sampled at 500 Hz, fused at 200 Hz.

Calibration and Mapping Models:

• Thermal–RGB Reprojection:

$$\mathbf{x}_{3D} = T_{cw}^{-1}\cdot[d\,K^{-1}[x\ y\ 1]^T]$$

where $K$ are the thermal camera intrinsics, $T_{cw}$ is the camera-to-world pose, and $d$ is ToF-provided depth.

• Tactile-to-Force Conversion:

$$F_{ij} = \alpha(v_{ij} - v_0) + \beta(v_{ij} - v_0)^2$$

with calibration over reference weights (20–500 g).

• IMU Bias Correction (EKF):

$$\mathbf{x} = [\mathbf{q}, \mathbf{b}_\omega, \mathbf{b}_a]^T$$

with discrete-time prediction and measurement updates for orientation and bias estimation.

This sensor array enables high-fidelity, real-time acquisition of force, visual, thermal, and kinematic data for advanced robotic manipulation (Zhou et al., 24 Jun 2025).

3. Electronics and Control Architecture

LEAP Hand V2 features a hierarchical, three-tier electronics and control stack:

• Tier 1 (Joint-Level): Each IMU communicates with a joint-dedicated STM32F405 MCU via RS485 (Modbus RTU, accel/gyro@500 Hz, mag@100 Hz).
• Tier 2 (Palm Aggregation): A custom PCB aggregates tactile, IMUs, and ToF sensor data. Multistream IMU fusion (EKF), tactile and ToF packaging, and communication with Tier 3 through UART (USB-C) interface.
• Tier 3 (Central Compute): Raspberry Pi 5 (8 cores @ 2.4 GHz, 4 GB RAM) operating ROS 2 for sensor synchronization, SLAM (SfM/dense recon), and grasp planning. It interfaces with the thermal camera (SPI + I²C), PiCam (CSI), ToF (I²C), and Tier 2 via UART.

Controller Hierarchy:

• Low-Level: Dynamixel torque: $$\tau = K_p(\theta_{des} - \theta_{act}) + K_d(\omega_{des} - \omega_{act})$$
• Mid-Level: Fuses IMU/tactile slip detection for pose/impedance control.
• High-Level: MuJoCo-supported Real2Sim pipeline for motion planning (Zhou et al., 24 Jun 2025).

4. Quantitative Performance Evaluation

Task-driven evaluations highlight the efficacy of LEAP Hand V2’s multimodal sensing:

• Temperature-Aware 3D Reconstruction and Safe Grasping:
  • 90 multi-modal (RGB, thermal, depth) captures processed via SfM and Gaussian splatting for meshing.
  • Thermal–RGB alignment as per the calibration equation.
  • Denoising removes 92% of spurious thermal outliers.
  • Imitation-learned grasp policy (5,000 frames/12 demonstrations) yields 95% success in avoiding regions $\pm15^\circ$0C, with zero contact failures over 30 trials.
• Mass Differentiation via “Flick”:
  • U-shaped objects (82 g, 125 g, 219 g) flicked using identical torques; 50 trials per mass.
  • 42 time-series features extracted (min/max/mean/std across axes, range, magnitude).
  • Linear Discriminant Analysis: LD1 captures 77.5% and LD2 22.5% variance; 3-class accuracy 96.4% (5-fold CV).
  • Dominant features: ACC_Z_Min, ACC_Range, MAG_Y_Mean.

These results substantiate the hand’s capacity for closed-loop, safe grasping and tactile-kinematic mass distinction unattainable by vision-only approaches (Zhou et al., 24 Jun 2025).

5. Reproducibility, Open-Source Resources, and Cost

LEAP Hand V2 emphasizes accessibility and community adoption:

Component	Quantity	Total USD
Dynamixel RX-10 servos	12	348
Tactile 6×6 arrays	5	500
BMM350 IMU modules	15	75
FLIR Lepton 3.5 camera	1	200
Raspberry Pi 5	1	80
Pi Camera Module 2	1	30
VL53L5 ToF depth sensor	1	50
PCBs/electronics	—	400
3D-printed PLA	—	200
TOTAL (≤$4k)		2,113

All design files (CAD, PCB layouts), MCU firmware, and ROS 2 software are open-source at https://github.com/slurm-lab-usc/motif-hand, including detailed assembly documentation, calibration routines, and demonstration scripts (Zhou et al., 24 Jun 2025).

6. Limitations and Prospective Advancements

LEAP Hand V2 exhibits several current limitations:

• Absence of dedicated fingertip GelSight/DIGIT sensors; tactile arrays are limited to the proximal phalanges.
• Native thermal camera resolution (160×120) constrains fine-grained heat mapping.
• IMU drift and high-frequency magnetic interference require periodic re-zeroing and shielding, respectively.

Proposed directions for future V3 development include:

• Integration of DIGIT 360 or GelSight at every fingertip for sub-millimeter contact geometry.
• Higher-resolution thermal sensing (e.g., Lepton 3.5 to Boson 320×256) with on-chip fusion.
• Addition of strain-gauge torque sensors at the actuator level.
• Soft-skin over-molding for human-like compliance, paralleling planned enhancements in the forthcoming LEAP Hand V2 by Shaw et al.
• End-to-end, visuo-thermal-tactile reinforcement learning for more adaptive manipulation policies.

The MOTIF/LEAP Hand V2, by coupling dense and multimodal sensing with an accessible open-source foundation, establishes a technically robust and reproducible platform for experimental research in dexterous robotic manipulation (Zhou et al., 24 Jun 2025).