CRAFT: A Tendon-Driven Hand with Hybrid Hard-Soft Compliance

The paper introduces CRAFT, a low-cost, open-source, tendon-driven anthropomorphic hand that combines rigid links with soft joints and rolling-contact surfaces to achieve robust, repeatable, and versatile contact-rich manipulation.

The Gist

Imagine you are trying to build a robot hand that is as dexterous as a human's but tough enough to survive being dropped, bumped, or used to pick up a fragile egg without crushing it.

For decades, engineers faced a "Goldilocks problem":

• Too Hard: Traditional robot hands are made of rigid plastic and metal. They are precise, but if they bump into something, they break, or they crush the object they are holding. It's like trying to pick up a grape with a pair of steel tongs.
• Too Soft: Soft robot hands are made of squishy rubber. They are great at hugging objects, but they are wobbly. They can't hold heavy things, and they are hard to control because they bend differently every time you squeeze them. It's like trying to write with a wet noodle.

Enter CRAFT: The "Best of Both Worlds" Hand.

The researchers behind CRAFT (a project from the University of Illinois and UC Irvine) came up with a clever solution: Don't make the whole hand soft or hard. Make it a hybrid.

The Core Idea: "Hard Bones, Soft Joints"

Think of CRAFT like a human hand.

• The Links (Bones): The long parts of the fingers are made of hard plastic (PLA). These act like your finger bones. They are rigid so they can carry heavy loads and move in a straight, predictable line.
• The Joints (Cartilage): The knuckles are made of soft, squishy rubber (TPU). These act like your cartilage. When the hand bumps into something, the rubber absorbs the shock, just like your knee absorbs the impact when you jump.

This design solves the main problem: The hand is strong enough to lift a dumbbell, but soft enough to pick up a potato chip without breaking it.

The Secret Sauce: Rolling Joints

Here is where the engineering gets really smart. Usually, if you make a joint out of soft rubber and bend it over and over, it gets tired and snaps (like bending a paperclip until it breaks).

CRAFT avoids this by using Rolling-Contact Joints.

• The Analogy: Imagine a door hinge. If you hinge a door, the metal rubs against metal in one spot. Eventually, it wears out. Now, imagine a wheel rolling on a track. The contact point keeps moving, so no single spot gets worn down.
• The Result: CRAFT's soft joints "roll" instead of "bend" at a single point. This spreads the stress out, making the hand incredibly durable. It can be used thousands of times without breaking.

How It Works: The Puppet Master

Instead of putting heavy motors inside the fingers (which makes the hand bulky and clumsy), CRAFT uses tendons.

• The Analogy: Think of a puppet. The strings (tendons) pull the fingers, but the puppeteer (the motors) is standing back in the forearm.
• The Benefit: This keeps the fingers light and thin, looking almost exactly like a human hand. It also means if a finger gets broken in a crash, you can just swap out that one finger without rebuilding the whole robot.

Why Does This Matter? (The "Real World" Test)

The team tested CRAFT against other robot hands, and the results were impressive:

1. The "Egg Test": When trying to pick up a raw egg or a raspberry, rigid robot hands often crush them because they can't "feel" the pressure. CRAFT's soft joints naturally squeeze just enough to hold the egg without breaking it.
2. The "Chopstick Test": Because the fingers can move side-to-side (a feature missing in many other cheap robot hands), CRAFT can pick up chopsticks or coins, which requires fine, lateral control.
3. The "Durability Test": In a test where they tried to pull the fingers apart, CRAFT held nearly twice as much weight as a standard rigid robot hand. The tendon system acts like a pulley, giving it extra strength.

The Big Picture

CRAFT isn't just a cool gadget; it's a tool for teaching robots how to think.

• For Data Collection: To teach a robot AI how to do things, humans need to show it by controlling the robot hand (teleoperation). With rigid hands, the human operator has to be perfect, or the robot will crash. With CRAFT, the hand is forgiving. If the human makes a slight mistake, the soft joints absorb the error, and the robot still learns successfully.
• Cost: It costs less than **$600** to build (compared to$15,000+ for high-end research hands), and the plans are free for anyone to download.

In summary: CRAFT is a robot hand that combines the strength of a steel bone with the flexibility of human skin. It's built to survive the messy, unpredictable real world, making it the perfect partner for teaching robots how to be truly dexterous.

Technical Summary

1. Problem Statement

The paper addresses a fundamental tension in robotic dexterous manipulation: the trade-off between precision and robustness.

• Rigid Hands: Traditional direct-drive hands offer precise kinematics and high load capacity but lack compliance. They transmit impacts directly to joints and linkages, making them prone to damage during inevitable collisions, slips, or suboptimal interactions in data collection.
• Soft Hands: Fully compliant hands absorb uncertainty passively but suffer from limited load capacity and configuration-dependent kinematics (motion varies based on load), making them difficult to model and control repeatably.
• The Gap: Existing solutions often apply uniform material properties (all rigid or all soft). The authors argue that manipulation requires a spatial separation of functions: joints (where impacts and contact uncertainty concentrate) need compliance, while links (which transmit actuation forces) need rigidity.

2. Methodology: The CRAFT Hand

The authors propose CRAFT, a tendon-driven, anthropomorphic hand that implements a hybrid hard-soft compliance strategy.

Design Architecture

• Material Distribution:
  • Links: Constructed from rigid Polylactic Acid (PLA) to ensure geometric predictability and load transmission.
  • Joints: Constructed from soft Thermoplastic Polyurethane (TPU) to provide passive compliance and impact absorption.
• Actuation:
  • Tendon-Driven: 15 motors are mounted on the forearm (not in the hand), driving the fingers via tendons. This keeps the hand lightweight (800g), compact, and anthropomorphic.
  • Degrees of Freedom (DoF): 15 active DoFs and 5 passive DoFs. Each finger has 3 actuators controlling: (1) coupled PIP/DIP flexion, (2) MCP abduction/adduction, and (3) MCP flexion/extension.
• Key Mechanical Innovations:
  1. Bidirectional Coupled Linkage: A mechanical linkage couples the Proximal Interphalangeal (PIP) and Distal Interphalangeal (DIP) joints. This ensures that regardless of external load, both joints bend equally and simultaneously, solving the unpredictability of soft tendon-driven fingers.
  2. Rolling-Contact Joints: Instead of traditional flexure joints (which fatigue and break at a fixed point), the PIP and DIP joints utilize rolling-contact surfaces. This distributes stress across the TPU surface, significantly reducing stress concentration and improving durability.
  3. Snap-Fit Interfaces: The MCP joints use a cylindrical snap-fit interface that "pops out" under excessive force, acting as a mechanical fuse to prevent permanent damage.

Software & Integration

• Teleoperation: Uses a single RGB camera with vision-based tracking (FrankMocap for arm, HaMeR for hand) to control the robot without gloves or wrist sensors. A calibration procedure maps human biological limits to the robot's joint limits.
• Simulation: Provides URDF and MuJoCo XML models. Soft-body physics are approximated using equality constraints and revolute joints to maintain computational efficiency while preserving kinematics for reinforcement learning.

3. Key Contributions

1. Hybrid Hard-Soft Design: A novel architecture that localizes compliance at joints and rigidity at links, resolving the precision-robustness trade-off at a low cost (<$600).
2. Coupled Rolling-Contact Joints: A mechanism that combines bidirectional linkage with rolling surfaces to produce predictable, repeatable motion while absorbing impact and preventing fatigue failure.
3. Open-Source Ecosystem: The full hardware design, simulation models, and teleoperation stack are released open-source to facilitate data-driven robot learning.
4. Comprehensive Validation: The hand is benchmarked against rigid baselines and validated across structural, manipulation, and dexterity metrics.

4. Results

A. Structural Performance

• Strength (Pull-out Test): CRAFT withstands 15.29 N of pull-out force, nearly double the 8.67 N of the rigid LEAP hand baseline. This is attributed to the mechanical advantage of the tendon routing.
• Repeatability: Despite soft joints, CRAFT maintains a joint angle tracking error of < 0.01 rad over one hour of cyclic loading, comparable to the rigid LEAP hand.
• Endurance (Holding Test): When holding a 5 lb weight for one hour, CRAFT consumes ~50% less current than the LEAP hand. The tendon friction acts as a passive brake, reducing motor strain.

B. Teleoperation & Manipulation

A user study compared CRAFT against the rigid LEAP hand across five tasks (Ball, Wine Glass, Egg, Raspberry, Frito Chip):

• Fragile/Low-Friction Objects: CRAFT achieved 100% success on delicate items (Raspberry, Chip) and 90% on the Egg, whereas the rigid hand failed frequently (60% failure rate) due to crushing or slipping.
• Efficiency: Operators using CRAFT completed tasks significantly faster (e.g., lifting a wine glass took half the time) because they could rely on passive compliance rather than precise force control.
• Dexterity: CRAFT successfully executed 33/33 grasp types in the Feix taxonomy, covering everything from power grips to precision pinches.

C. Comparison to State-of-the-Art

• vs. RUKA & ORCA: Unlike RUKA (lacks abduction) and ORCA (fixed DIP), CRAFT supports full lateral movement and active fingertip control, enabling tasks like holding chopsticks.
• vs. Leap V2: CRAFT is significantly more compact and affordable ($600 vs.$5,000). Its slim profile allows it to access narrow spaces (e.g., inside a jar) where the bulkier Leap V2 fails.

5. Significance

The CRAFT hand represents a significant step forward in making dexterous manipulation accessible for robot learning.

• Data Collection: By reducing the risk of hardware failure during collisions and lowering the cognitive load on human operators, CRAFT enables the collection of large-scale, high-quality manipulation datasets.
• Scalability: The low cost and open-source nature allow researchers to deploy multiple units for parallel data collection.
• Robustness: The hybrid design proves that high dexterity does not require sacrificing durability; the hand can survive the "reality" of learning (lots of contact and repetition) while maintaining the kinematic structure necessary for precise control.