The Cartesian Hand replaces finger joints with seven linear axes

Duke University's General Robotics Lab introduced the Cartesian Hand in a preprint submitted on September 22, 2026. Instead of imitating an articulated human hand, the 850-gram end effector stacks two independently actuated parallel grippers and lets four fingertips translate along fixed axes. The seven-degree-of-freedom mechanism targets objects with threads, pivots, guides, plungers and triggers. Paper and version record.

Key takeaways

• All seven joints are prismatic: four fingertip slides, two gripper openings and one axis that changes the separation between the grippers.
• The authors report 350 successful trials across 35 configured objects, with 10 trials per object. Objects began from prescribed positions, and new objects within an established category could require up to five setup trials.
• The hand uses joint feedback and contact detection but does not directly measure contact location or contact force. Its strength comes from mechanism-specific motion, not general tactile dexterity. Methods and tests.

What changed

Most dexterous hands add rotary joints so fingers can wrap around objects. The Cartesian Hand assigns separate roles to two simple grippers. One section can hold the body of a bottle, pipette or tool while the other moves a cap, plunger, second handle or trigger. Because every joint translates along a fixed axis, the fingertip position is a linear function of joint position and the Jacobian does not change with configuration.

Seven Feetech-3915 servos drive rack-and-pinion stages and dovetail slides. The paper reports 65 millimeters of fingertip travel, 52 millimeters for each gripper opening and inter-gripper separation, and joint speeds around 60 millimeters per second. The closed assembly is approximately 166 by 100 by 76 millimeters. One gripper held a 2-kilogram bottle in a static test.

The authors estimate about USD 500 for a complete hand. Structural printing is estimated at roughly USD 30 in PLA or USD 50 in SLS Nylon 12, and PLA assembly takes about two hours after printing. These are author estimates, not independently audited production costs. A theoretical rack force near 170 N derives from rated servo stall torque before transmission losses; it is not a measured continuous fingertip force.

What the 350 trials establish

The evaluation covers cap opening and closing, pipetting, pumping, two-handle tools, screwdrivers, triggers and in-grasp reorientation. Ten trials on each of 35 objects all succeeded after object-specific parameters were configured. During the in-hand procedure, the Franka Panda arm remains stationary and an external holder no longer supports the object after lift. Experimental setup and results.

That repeatability is meaningful for a structured mechanism, but it is not zero-shot generalization. Initial object poses are prescribed. A new object in an existing mechanism category reuses the procedure, then changes settings such as grasp height, stroke or force limit; the paper allows up to five setup trials to find them.

The project also shows two Cartesian Hands on Duke Humanoid V2 uncapping a tube, pipetting and recapping. This is a qualitative transfer demonstration. The paper does not report a second 350-trial evaluation on the humanoid. Paper's humanoid demonstration.

What this means for robotics

RoboSkin analysis: dexterity does not always require anthropomorphic fingers. Many industrial and laboratory objects already constrain motion through threads, hinges and guides. Designing a hand around those mechanisms can simplify planning and make a small library of primitives reusable. The result belongs alongside underactuated and anthropomorphic designs in any robot hand comparison.

It also highlights where touch would add value. Joint thresholds can detect resistance, but they cannot identify which surface made contact or separate desired tool load from a collision. Adding compact tactile sensors is a possible route to observing those missing contact variables. Whether it enables automatic parameter selection, slip response or better force limiting would require a new integrated evaluation; it is not a capability demonstrated here.

Limitations and availability

This is an arXiv v1 preprint and RoboSkin.ai has not built the hand. The object set is structured, the starting poses are controlled, and the procedures use manually selected parameters. The platform does not observe object pose, contact location or contact force. Claims should therefore remain about repeatable mechanism-specific manipulation, not open-world hand autonomy.

Code, CAD and build resources

The repository that was unavailable at our September 23 check is now public. On October 17, we verified revision 83b738e, committed October 10, and its root Apache-2.0 license. Use these versioned resources when evaluating a build:

The release does not independently reproduce the paper’s 350 trials. Its README explicitly describes a rewritten control stack, not the software used for those results, and says the transcribed manipulation tasks have not been completed on their physical objects. The mock and MuJoCo examples also contain no objects to manipulate. A completed simulation program therefore does not establish a successful grasp. RoboSkin.ai has inspected the public files but has not run this software or built the hand.

For evaluating contact feedback in a future build, use the robot-hand tactile sensor selection guide; the released hand should not be described as having that added tactile capability.

Linear finger motion does not make object control linear

The constant Jacobian relates seven joint displacements to fingertip reference positions. Turning a cap still depends on where the fingers engage, whether they slip, and how the cap constrains motion. Thus a simpler kinematic model reduces one planning problem without solving contact estimation. A useful receiving test separates these layers. Kinematics and manipulation methods.

The paper's detector recognizes sustained low joint velocity while movement is commanded. It can estimate a width or diameter from closure position, but that estimate depends on contact geometry and deformation. It is not a force sensor. Likewise, opening a cap or moving a plunger does not itself establish metering accuracy.

The November 4 recheck of the pinned hardware notes also matters: software travel limits, encoder scaling and zero references require build-specific verification, and the released task stack is distinct from the paper's implementation. Keep successful program execution separate from successful physical manipulation. Use the robot-hand evidence comparison to connect a build's mechanics, observations and completed task rather than transferring the paper's 350/350 result to an untested assembly.

Sources

• arXiv v1 record, submitted September 22, 2026
• Full Cartesian Hand v1 paper
• Official General Robotics Lab project page

*Source review updated November 4, 2026. Original publication dates are retained. Research results are author-reported; RoboSkin has not built or run these systems.*