Overview
Dexterous manipulation is the frontier of robot learning. While parallel-jaw grippers handle 80% of industrial pick-and-place, the remaining 20% -- tool use, in-hand rotation, assembly of flexible objects, opening containers -- requires multi-fingered hands with 12+ degrees of freedom. The choice of hand determines your research trajectory: actuation type affects control bandwidth, tactile sensing affects data richness, and price affects how many hands you can afford to break during experiments.
Full Specification Comparison
| Hand | DOF | Actuation | Tactile | ROS2 | Weight | Price |
|---|---|---|---|---|---|---|
| Allegro Hand v4 | 16 | Direct drive (Maxon motors) | Optional (add-on BioTac or custom) | Yes (allegro_hand_ros2) | 1.1 kg | ~$15,000 |
| LEAP Hand | 16 | Direct drive (Dynamixel servos) | None built-in | Community (Python SDK) | 0.5 kg | ~$2,000 (DIY) |
| Shadow Dexterous Hand E | 24 | Tendon-driven (pneumatic muscles or motors) | BioTac SP (optional, 19-taxel) | Yes (sr_ros2_interface) | 4.0 kg (with actuator pack) | ~$120,000 |
| Orca Hand | 12 | Tendon-driven (brushless motors) | Optional Paxini integration | Yes | 0.6 kg | Contact SVRC |
| Inspire Hands (RH56DFX) | 6 (1 per finger) | Underactuated (motor per finger) | None | Python SDK, Modbus RTU | 0.5 kg | ~$3,000 |
| RH8D (Seed Robotics) | 8 | Direct drive (Dynamixel) | Optional FSR arrays | Yes | 0.58 kg | ~$8,000 |
| Paxini Tactile Glove | N/A (sensor) | N/A | High-resolution capacitive array, full-hand coverage | Yes (ROS2 topic) | 0.12 kg | Contact SVRC |
Actuation Types Explained
Direct Drive
Motors are located in or near each joint. Allegro Hand uses Maxon DC motors; LEAP Hand uses Dynamixel smart servos. Advantages: simple kinematics, direct position/velocity/torque control, no cable routing. Disadvantages: motors add bulk and weight to the hand; limited force output for hand-sized motors.
Tendon-Driven
Motors are located in the forearm or a remote actuator pack, connected to joints via tendons (cables or rods). Shadow Hand and Orca Hand use this approach. Advantages: lightweight fingers, higher force (motors can be larger), more human-like mechanical design. Disadvantages: complex cable routing, hysteresis from cable stretch, more difficult to model accurately for sim-to-real transfer.
Underactuated
Fewer motors than joints. Inspire Hands use one motor per finger to drive multiple phalanges through a linkage mechanism. The finger adapts to object shape passively. Advantages: very simple control, robust, low cost. Disadvantages: cannot independently control each joint, limited dexterity for in-hand manipulation.
Tactile Sensing Integration
Tactile sensing is increasingly important for dexterous manipulation research. Adding tactile data to visual policies improves grasp success rates by 15-30% in recent studies. Options:
- Paxini sensors (available at SVRC): High-resolution capacitive tactile arrays that can be integrated with Orca Hand and other hands. Provide rich contact geometry data over ROS2 topics. Best option for data collection research.
- BioTac SP: 19-taxel fingertip sensor by SynTouch. Previously the gold standard but increasingly hard to source. Works with Shadow Hand and can be adapted to Allegro.
- GelSight / DIGIT: Vision-based tactile sensors that produce high-resolution contact images. GelSight Mini (~$300 DIY) and DIGIT (~$500) can be mounted as fingertips on custom hand designs. See our Tactile Sensor Comparison guide.
- FSR arrays: Lowest cost option. Force-sensitive resistors ($5-$50) provide binary or low-resolution pressure readings. Suitable for detecting contact presence but not for fine manipulation research.
Recommendations by Use Case
- Best budget research hand: LEAP Hand (~$2,000 DIY). Open-source design, 16-DOF, large community. Requires assembly and Dynamixel servo purchasing.
- Best for sim-to-real transfer: Allegro Hand v4 (~$15,000). Most widely used in RL-for-dexterity research. Accurate sim models in MuJoCo and Isaac Sim. Direct drive simplifies modeling.
- Best with tactile sensing: Orca Hand + Paxini sensors (available from SVRC). Purpose-built integration for tactile data collection.
- Highest dexterity: Shadow Dexterous Hand (~$120,000). 24-DOF, closest to human kinematics. Used by OpenAI (Rubik's cube), DeepMind, and top research labs.
- Best for humanoid integration: Inspire Hands (~$3,000). Lightweight, affordable, robust. Used on Unitree G1 and other humanoid platforms.
- Best for data collection suitability: Orca Hand or Allegro Hand with Paxini sensors. The combination of controllable fingers + rich tactile data produces the highest-quality manipulation datasets.