The Cartesian Hand: In-Hand Manipulation with All-Linear Fingers

TL;DR

The Cartesian Hand is a 7-DoF end-effector from Duke’s General Robotics Lab in which every actuated joint is prismatic — two independently actuated parallel grippers plus four translating fingertips, no rotary articulation anywhere. Because all joints are linear, the fingertip Jacobian is a constant full-rank matrix everywhere in the workspace: no singularities, no direction-dependent behavior, and manipulation decomposes into linear motion primitives. The authors demonstrate cap opening/closing, pipetting, pumping, screwdriving, trigger actuation, two-handle tools, and in-grasp reorientation across 35 objects (lab, manufacturing, household), plus a Rubik’s cube with the arm held still, and transfer the same procedures to a humanoid using two hands. The complete design is open-sourced at ~$500 in parts.

Key claims

Seven prismatic DoF (two 1-DoF parallel grippers + four translating fingertips) give a configuration-independent, constant fingertip Jacobian, eliminating singular poses and direction-dependent behavior across the workspace [§Abstract, Fig. 1].
Manipulation composes from linear motion primitives — no rotary articulation is needed to operate objects whose internal mechanism already constrains relative motion (threads follow a helix, plungers translate, triggers move along a fixed path) [§Abstract].
Objects that share a mechanism share a procedure: a thread is a thread whether it belongs to a centrifuge tube or a peanut-butter jar, so switching objects only requires updating grasp height, stroke, and force limit rather than rewriting the controller [Project description].
Demonstrations span 35 objects across lab (pipetting, centrifuge caps), manufacturing (screwdriving, trigger tools), and household (bottle caps, two-handle tools), including in-grasp reorientation and a Rubik’s cube solved with the arm held still [§Abstract].
The same manipulation procedures transfer from a fixed-base arm to a humanoid with one Cartesian Hand on each arm, demonstrating bimanual laboratory manipulation (uncap tube, pipette, recap) [§Abstract].
Hardware and software will be open-sourced; the hand can be 3D-printed and the parts cost is ~$500 [§Abstract, project page].

Method

Two independently actuated parallel grippers each provide one DoF of jaw motion. Four additional prismatic fingertips translate along fixed axes to generate relative motion between whatever the two grippers happen to be holding. Because every joint is linear, the map from joint positions to fingertip positions is a fixed linear function — the Jacobian is constant, so any inverse-kinematics or force computation reduces to a matrix that never has to be re-evaluated at runtime and never approaches singularity. The task interface exploits this: manipulation procedures are authored as compositions of linear primitives (translate this fingertip by Δx while holding gripper A, then release, then translate gripper B by Δy, …) rather than as full-arm trajectories in end-effector SE(3). The control stack (task authoring + tensor engine + three backends — hardware, CPU MuJoCo, batched GPU MuJoCo via NVIDIA Warp) is open-sourced separately from the mechanical design.

Results

35 objects across three domains (lab / manufacturing / household), including cap open/close, pipetting, pumping, screwdriving, trigger actuation, two-handle tools, and in-grasp reorientation [§Abstract].
Rubik’s cube solved with the robot arm held still — the hand alone provides enough relative motion between the two grasped halves to twist faces [Project description].
Zero-shot procedure transfer from arm to humanoid: bimanual uncap–pipette–recap demonstrated with two Cartesian Hands on a humanoid, using the same procedures authored for the arm [§Abstract].
Hardware cost: ~$500 for the complete hand, 3D-printable, open-sourced [§Abstract, Project description].

Why it’s interesting

The Cartesian Hand stakes out a genuinely new position on the open-source-dexterous-hand design axis relative to the two hands already filed. ORCA: An Open-Source, Reliable, Cost-Effective, Anthropomorphic Robotic Hand for Uninterrupted Dexterous Task Learning (ETH ORCA, 17-DoF, ~2,000 CHF) and Aero Hand Open: low-cost, open-source, lightweight, anthropomorphic dexterous hand (Aero Hand Open, 16 joints / 7 motors, $314) both remain anthropomorphic and both invest their DoF budget in *rotary* finger articulation to mimic the human hand — ORCA at the high-DoF-per-motor end and Aero Hand Open at the coupled-tendon low-motor end. The Cartesian Hand does the opposite: it *gives up* rotational articulation entirely and buys back tractability (constant Jacobian, no singularities, linear-primitive control) plus$ 500 parts cost. This is the same trade-off logic that RLWRLD’s How we actually test a robot hand hand-testing series argues for from the reliability side — simpler kinematics mean fewer parts to break — but from the control side rather than the mechanical side. It also creates an interesting contrast with the retargeting-heavy branch of the field: papers like Do as I Do: Dexterous Manipulation Data from Everyday Human Videos are trying to close the human-video → robot-hand gap through better retargeting, whereas the Cartesian Hand argues the field may not need an anthropomorphic hand at all if the task is “operate objects whose mechanism already constrains motion.” Whether that hypothesis holds outside the paper’s 35-object set — especially for tasks that require genuine rotational in-hand primitives like wrapping cloth or braiding wire — is the open question.