Recreating CRAFT: Hardware Lessons for Dexterous Manipulation

Aditya Mittal — Tue, 23 Jun 2026 14:50:27 -0400

Rebuilding CRAFT from the hardware up: compliant fingers, tendon routing, robot mounting, Quest-based teleoperation, and early policy rollouts.

Completed CRAFT build: compliant fingers, routed tendons, compact actuators.

CRAFT mounted on the Yam arm in the bimanual workspace.

Vision teleoperation: orange pick-and-place.

Most robot grippers are effective clamps: they open, close, and work well when the object is already in the right place. Dexterous hands support more complex forms of manipulation: rotating objects in-hand, wrapping around fragile items, using tools, and recovering grasps when contact is imperfect. This post documents my experience building the CRAFT hand at the UNVEIL Lab, originally developed by Leo Lin et al. at UIUC. The design principle is simple and appealing: passive compliance helps the hand adapt to contact, while tendon-driven actuation keeps the hardware compact.

The Original CRAFT Hand

CRAFT stands out because it combines low-cost hardware with passive compliance. The hand costs under $600, weighs 800 g, and can be built mostly from 3D-printed parts and off-the-shelf components. It uses rigid PLA where the hand needs structure and soft TPU in the finger joints, where impacts and contact uncertainty matter most. This makes the hand more forgiving around fragile, deformable, or irregularly shaped objects without turning the whole finger into a soft, hard-to-control appendage.

The hand remains dexterous: each finger has base flexion/extension, base abduction/adduction, and coupled middle/fingertip flexion, giving the hand 15 active degrees of freedom plus 5 passive degrees of freedom. The motors sit away from the fingers and pull through tendons, which keeps the fingers light and close to human scale. The original paper reports coverage of all 33 Feix grasp types and demonstrates teleoperation with delicate objects such as eggs, raspberries, chips, and wine glasses.

Original CRAFT prototype from the project paper, showing the compact tendon-driven layout and hybrid rigid-soft fingers.

Step 1: Building the Compliant Fingers

The build started with the compliant fingers. Each finger combines white PLA link shells with black TPU joint pieces. The PLA provides structure and clean mounting interfaces, while the TPU forms the bending regions so the finger can deflect during contact instead of behaving like a fully rigid linkage. During assembly, the main constraint was alignment: the pins, fasteners, and printed interfaces had to seat cleanly without compressing the TPU enough to restrict joint motion.

I first built a single finger to validate the PIP/DIP stack before repeating the process across all five fingers. The finger needed to bend smoothly by hand, return without obvious binding, and avoid side-loading between the printed parts. Once that validation passed, I assembled the remaining finger modules and staged them around the palm for tendon integration.

Single finger module: PLA forms the rigid links, while TPU adds compliance at the joints.

Finger modules during assembly. The flexible TPU joints made alignment important before tendon routing.

Finger modules staged around the PLA palm plate before tendon routing and final mounting.

Tendon routing turned the finger modules into actuated fingers. Pulling the flexion tendon curls the finger, while the rolling-contact geometry keeps the PIP and DIP joints moving together instead of letting the fingertip fold independently. The key requirement was not just motion, but repeatable force transmission. A small rubbing point or sharp exit angle can produce weak, delayed, or inconsistent fingertip motion from the same motor command.

Manual tendon test during assembly, checking side-to-side motion, finger bend, and curl before motor integration.

First desktop teleoperation test, mapping tracked hand motion to CRAFT motor commands.

Step 2: Mounting CRAFT on the Existing Bimanual Setup

After bench assembly, the next step was integrating CRAFT with the existing bimanual I2RT Yam Ultra setup. The original setup used stock parallel-jaw grippers, which are reliable for simple pick-and-place tasks but reduce the end-effector command to opening and closing. Mounting CRAFT on the wrist turned the hand from a standalone build into part of the robot platform. The payload budget was sufficient: the Yam Ultra is rated for 4 kg, while CRAFT is reported at about 800 g, leaving margin for the adapter, fasteners, and wiring.

Yam Ultra arm with the stock parallel-jaw gripper before the CRAFT retrofit.

To adapt the arm, I used Autodesk Fusion and the available I2RT STL files as the starting geometry for a printable wrist mount. The design requirements were mostly mechanical: match the Yam wrist pattern, interface cleanly with the CRAFT mount, keep the fingers oriented toward the workspace, preserve wrist clearance, and leave a usable path for wiring. I checked the wrist motion in MuJoCo before printing, then mounted the hand on the left arm and staged the workspace for teleoperation.

CRAFT mounted on the left Yam arm, with the bimanual workspace staged for teleoperation.

Top view of the mounted hand, showing its position relative to the tabletop objects.

Step 3: Quest-Based VR Teleoperation

With the hand mounted, the next requirement was a control interface for collecting demonstrations. We used Quest-based VR teleoperation: the headset tracks the operator's hand pose, and the retargeting layer maps that motion into CRAFT's motor limits. The thumb mapping required additional calibration because small tracking or frame errors produced large changes in the commanded thumb motion.

For arm motion, we used an Open-TeleVision-style setup in which the operator's wrist pose commands the robot wrist and guides the arm. When the wrist frame was aligned, the arm and hand could be controlled as one coordinated system rather than as two separate interfaces. The early demonstrations included simple pick-and-place tasks, followed by a wine-pouring setup assisted by GLIDE guardrails. In that setting, GLIDE filters unstable commands around the glass grasp and pour while the operator still provides the task-level motion.

Quest-based VR teleoperation with CRAFT mounted on the robot wrist.

Pick-and-place with a white cup, where compliance helps the fingers conform without over-squeezing the object.

More complex wine-pouring setup with GLIDE guardrails, where the hand, robot, camera, and workspace must stay coordinated.

Pi0.5 Policy Rollout

After collecting human demonstrations, we tested a rollout using pi0.5 with GLIDE assistance. This setup changes the role of the operator: instead of directly commanding every motion, the learned policy attempts the wine-serving behavior while GLIDE filters unstable commands around the grasp and pour.

Pi0.5 wine-serving policy rollout with GLIDE guardrails filtering unstable commands.

Challenges and Lessons Learned

As a beginner, some lessons were obvious only after the fact. The hard part was turning a printed hand into a dependable robot system: parts, assembly, calibration, contact, and real-robot debugging.

Spares matter more than expected. The parts list was not the issue; my assumption was that exact quantities would be enough. A broken motor, missing screw, or failed print can turn into a one- or two-week delay without extra screws, bearings, tendons, inserts, and at least one spare actuator.
Retargeting was not plug-and-play. The Quest gave us a usable hand pose, but the mapping to CRAFT still needed calibration. Thumb motion needed more useful side-to-side control, and even a small wrist-frame mismatch could make the robot feel unintuitive.
Contact was more than closing the fingers. The hand could wrap around a bottle and still slip during pouring, which made it clear that enclosing an object is not the same as gripping it. Friction tape helped temporarily, but better fingertip pads or coatings would make contact more reliable.
Motor torque became a real hardware limit. The small DYNAMIXEL servos keep CRAFT compact and affordable, but they limit grasp strength during stronger holds, slip recovery, and tasks where the object applies a moment to the fingers.
Simulation stopped short of the full system. Once CRAFT was mounted on the Yam arm with the adapter and tendon routing, many issues only appeared on hardware. A better Yam + CRAFT simulation would make it easier to test retargeting, wrist alignment, and policy rollouts before real-robot trials.

Next Steps

The next goal is to turn the build from a working prototype into a repeatable experiment platform. That means reducing the amount of manual adjustment needed between runs and making failures easier to diagnose.

Standardize maintenance. Keep a small spare-parts kit for screws, bearings, tendons, inserts, and actuators, then document the assembly order and common failure points so repairs do not depend on memory.
Make tendon routing more repeatable. Cleaner exits, better retensioning points, and less rubbing would make finger motion more consistent. Finger modules should also be easier to remove without disturbing the rest of the hand.
Improve fingertip contact. Replace the temporary friction tape with pads or coatings that provide repeatable grip without making the fingers too stiff. This would make pouring and heavier grasps less sensitive to slip.
Turn calibration into a procedure. Thumb retargeting and wrist-frame alignment should use a short routine with saved parameters, rather than being adjusted by feel each session.
Build an integrated Yam + CRAFT simulation. The simulation should include the adapter, wrist limits, approximate hand geometry, and task objects, so retargeting and simple policy changes can be tested before hardware trials.
Add contact feedback. Fingertip tactile sensing, or even simpler slip/contact indicators, would make it easier to understand why a grasp fails instead of relying only on camera views and visual inspection.

Aditya Mittal - About Me