An open robotic forearm tests how eight carpal bones redirect wrist stiffness

Researchers at the University of Electro-Communications in Tokyo released an anthropomimetic soft robotic forearm on September 24, 2026, to test how carpal-bone structure changes wrist stiffness. The prototype reproduces eight independently movable carpal bones, 22 actuated muscles, 13 finger degrees of freedom and a three-degree-of-freedom wrist. Its open companion repository includes CAD, printable parts, firmware, measured data and analysis code. Paper and version record.

Key takeaways

• The study compares one anatomical wrist, a fused proximal carpal row and a geometric ellipsoidal skeleton under four muscle-activation patterns.
• With the anatomical skeleton, switching from finger-only to combined wrist-and-finger activation rotates the stiffness ellipse by 48.2 degrees; a permutation test reports p equals 0.0038.
• The linked release includes mechanical, control and analysis materials, but licenses differ by file: derived bone models and fingertip molds use CC BY-SA 2.1 JP, while the authors' other materials use CC BY 4.0. Repository and license map.

What changed

Many robot wrists simplify the human carpus into one or two joints. This forearm instead uses CT-derived bone geometry, knitted polyethylene ligaments, tendon sheaths, a printed triangular fibrocartilage complex and dual-layer silicone fingertips. Nineteen muscles use tendon transmission; three intrinsic thumb muscles enable opposition. The assembled hand can grasp while its wrist moves, but the research question is mechanical rather than task-level autonomy.

To isolate morphology, every active muscle receives the same 0.6 N tension. A force gauge deflects the wrist 3 mm from 12 directions at 30-degree intervals, while motion capture samples at 100 Hz. The team runs eight trials per activation condition and direction, fits a stiffness ellipse, then separately records 20 repetitions per condition of motion between selected carpal bones.

That protocol matters when interpreting the result: the paper tests whether structure can redirect stiffness under uniform inputs. It does not optimize tendon forces or show an autonomous robot hand choosing stiffness for a manipulation task.

What the stiffness experiment found

For the anatomical skeleton, finger-muscle activation places the low-stiffness axis near the human dart-throwing direction. Combined wrist-and-finger activation instead aligns the high-stiffness axis with that direction. The observed 48.2-degree major-axis shift exceeded 4,982 of 5,000 shuffled-label outcomes.

Fusing the proximal row changes that behavior. Under finger activation, minimum-axis stiffness rises from 80.1 N/m for the anatomical skeleton to 149.5 N/m for the fused skeleton, and the low-stiffness direction moves away from the dart-throwing range. The ellipsoidal skeleton is stiffer overall, reaching major-axis values from 282.2 to 518.3 N/m under finger-only and wrist-only activation, but its ellipse orientation stays between about 86 and 96 degrees across conditions.

The carpal-motion test provides a narrower mechanistic result. Activation condition significantly changes relative rotation and translation at the proximal carpal row, while the midcarpal changes do not reach significance. The paper treats the latter as a supporting trend rather than a confirmed effect. Stiffness and motion results.

What this means for robotics

RoboSkin analysis: high stiffness alone is not the same as useful stiffness. The ellipsoidal wrist has the highest measured major-axis stiffness across the tested activation conditions, but changes its stiffness direction little. Stiffness is resistance per displacement, not a destructive strength or payload rating. For contact-rich manipulation, morphology that changes where the hand yields may reduce how much active control is needed before tactile feedback reacts.

The open hardware is also valuable as a reproducibility package. Engineers can inspect the bone and jig CAD, bill of materials, Dynamixel firmware and analysis scripts instead of inferring construction from figures. Those artifacts support a more specific reproduction plan, although their presence alone does not establish that a second build will reproduce the result.

Limitations and availability

All stiffness measurements are quasi-static. The paper does not evaluate impacts, fast grasp corrections, hammering, autonomous manipulation success or repeated builds. Every reported result comes from one physical prototype. Assembly jigs and current control improve consistency within that build, but cannot establish inter-build variation.

The repository was publicly accessible on September 25 and contained analysis data, Python code, PlatformIO firmware, CAD, STL files, a bill of materials and video. Multiple licenses apply, so “open hardware” does not mean every file has the same reuse terms. Users must preserve the share-alike license for the derived BodyParts3D components where applicable. RoboSkin.ai did not fabricate or test the design.

What a second build would need to reproduce

The experiment asks whether morphology redirects resistance under controlled tendon activation. It does not show an online controller selecting the best stiffness for each task. The 100 Hz motion-capture stream measures the experiment; it is not a demonstrated adaptation rate. Methods and supplementary assembly notes.

Three records would make a reproduction interpretable:

1. Mechanical state: record bone revision, ligament attachment and resting length, tendon route, fingertip construction and neutral pose. The authors' jigs control several of these variables; a nominally identical CAD model alone does not.

2. Measurement and exclusions: preserve loading direction, contact reference, force/displacement synchronization and every repeated measurement. The analysis excludes negative stiffness and values below 5 N/m for stated measurement reasons. Report exclusions instead of presenting the retained ellipse as all observations.

3. Repeatability level: distinguish repeated pushes on one assembly from rebuilding the assembly. The permutation result addresses within-build variation; dynamic impacts and variation among separately built forearms remain different tests.

For a task-specific design decision, ask which direction should yield and which should resist, then evaluate that choice under the intended motion. A large stiffness value can be undesirable along a direction that must accommodate misalignment. This is an engineering interpretation, not a manipulation-performance result from the paper.

As reviewed on November 4, the versioned release retains separate licenses for anatomy-derived parts and the specified original materials. Check the file map alongside the robot-hand mechanics comparison; no analysis scripts or hardware were executed for this review.

Sources and related resources

• arXiv v1, submitted September 24, 2026
• Full paper and supplementary material
• Official open-hardware repository
• RoboSkin soft robotic skin overview

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