Two-DoF distal wrist · ICRA 2027 manuscript

DistalShift

How distal redundancy shapes robot arm motion and joint loading.

Tian-Ao Ren^1*, Tongyu Wang^1,2*, Shixuan Zhao¹, Rohan Karunaratne¹, Tao Zhang¹, Rachel Thomasson¹, Mark Cutkosky¹

¹ Department of Mechanical Engineering, Stanford University · ² Department of Modern Mechanics, University of Science and Technology of China
* Equal contribution · tianao@stanford.edu, tongyuu@stanford.edu, tongyuw@mail.ustc.edu.cn

Paper PDF Watch the video BibTeX

44.5–95.8% lower reconstructed incremental proximal swept volume across five tasks

18.9–89.4% lower baseline-subtracted proximal torque excursion

48.4% / 29.8% lower peak and raw RMS arm-joint torque during contact roll

Abstract

Seven-degree-of-freedom manipulators provide redundancy for six-dimensional tool-pose control, yet local tool reorientation can still require substantial proximal-arm motion.

We present a compact two-degree-of-freedom distal wrist that localizes tool reorientation near the endpoint. The Contact-Anchored Reorientation Envelope evaluates orientation coverage under a proximal-motion budget and supports comparisons of mounting and arrival strategies. Hardware experiments compare wrist-held and wrist-actuated configurations for ultrasound probe manipulation across five free-space and phantom-contact conditions, retaining the same installed hardware.

Wrist activation reduced mean reconstructed incremental proximal swept volume by 44.5–95.8% and baseline-subtracted torque excursion by 18.9–89.4%. During axial probe rotation in phantom contact, peak and RMS measured arm-joint torque decreased by 48.4% and 29.8%, respectively. These gains were task dependent and accompanied by increased orientation-tracking error. Complementary diagnostic simulations indicate lower whole-system absolute joint mechanical work than direct probe mounting, including the wrist joints’ contribution, despite higher proximal torque RMS.

Overview

DistalShift adds local tilt and axial rotation between a commercial 7-DoF arm and the tool. The same assembly is compared with the wrist held, so the proximal arm supplies the reorientation, and with the wrist actuated, so more of that reorientation stays near the probe.

Contact roll and contact pitch illustrations beside swept-volume and torque-excursion charts for locked 7-DoF and active 9-DoF wrists. — Contact probe reorientation with a locked 7-DoF or active 9-DoF distal wrist. The charts show incremental proximal swept volume reconstructed from hardware trajectories and baseline-subtracted torque excursion. Bars are means; error bars are sample standard deviations, with n = 3 per condition.

Video

The project video walks through the 7-DoF and 9-DoF comparison, first in simulation and then on the hardware used for the ultrasound-probe tasks.

DistalShift project video.

Hardware

The wrist is an add-on module between the robot flange and the tool. Joint 8 tilts the tool and joint 9 rotates it about the probe axis. In the ultrasound configuration, that axial joint is aligned with the acoustic axis. All reported hardware experiments used the transverse 90° mount. The arm, probe, and phantom stay the same when the wrist is held or actuated.

Arm: Flexiv Rizon 4s, seven proximal joints
Wrist: HighTorque HTDW-5047-36-NE and HTDW-4438-32-NE, 36:1 and 32:1
Ranges: ±90° at joint 8 and ±180° at joint 9
Mass: 720 g excluding the probe
Tool offset: 290.4 mm from flange to the nominal contact point
Materials: Markforged Onyx base and mounts; Formlabs Rigid 4000 bracket, pulley, and connector
Probe: Butterfly iQ+ on an ABDFAN ultrasound-training phantom
Rates: Local wrist control at 500 Hz; host references and logging near 100 Hz

Interactive model

Drag a link to move its joint, or use the sliders. Drag empty space to orbit, and scroll or pinch to zoom. The viewer uses the original Onshape URDF, with the same joint axes as the hardware.

Loading interactive wrist model…

Drag a movable part to actuate its joint Drag empty space to orbit · wheel or pinch to zoom

Locked wrist during a pitch motion over the ultrasound phantom. — Hardware stills from pitch and axial-rotation sequences. With the wrist active, the proximal arm remains closer to its starting posture while the tool changes orientation.

Active wrist during a pitch motion over the ultrasound phantom. — Hardware stills from pitch and axial-rotation sequences. With the wrist active, the proximal arm remains closer to its starting posture while the tool changes orientation.

Hardware recordings

Each pair shows the same class of scripted probe motion with the wrist held and with the wrist actuated. Free-space pitch and roll leave the probe off the phantom. Contact pitch keeps the probe against the phantom while the tilt follows the reference.

Free-space pitch · locked

Free-space pitch · active

Free-space roll · locked

Free-space roll · active

Contact pitch · locked

Contact pitch · active

Simulation

The same probe motion is also compared in simulation. The 9-DoF model keeps the positioning arm quieter while the distal wrist supplies the orientation change.

Simulation · 7-DoF

Simulation · 9-DoF

Reorientation coverage

The Contact-Anchored Reorientation Envelope asks which probe orientations are reachable at a fixed torso contact when proximal joint motion, joint limits, clearance, and pose tolerance are bounded. It is a kinematic capability check, not a measure of force tracking or continuous trajectory feasibility.

On the 0° assembly, sampled coverage was 6.74%, 29.90%, 71.67%, and 88.50% for locked 7-DoF, generic 9-DoF, distal-weighted 9-DoF, and the proposed arrival-aware strategy. On the 90° assembly the corresponding values were 5.15%, 34.45%, 87.21%, and 94.96%. Coverage therefore depends on how the extra joints are used and how the arm arrives, not only on adding two degrees of freedom. Proposed-method coverage was 6.46 percentage points higher for the 90° assembly.

CARE comparison of 0 degree and 90 degree mounts, including coverage heatmaps and bar charts for four inverse-kinematics methods. — CARE comparison of the original 0° and transverse 90° mounts. Heatmaps use a common 0–100% scale. Archived seeds and hardware revisions differ, so this is a complete-assembly comparison rather than an isolated mount-angle ablation.

Hardware results

Five scripted conditions compare the locked and active wrist on the same installed hardware, with three technical repetitions each. Free-space tasks are an axial 0°→90°→0° turn and a 100 mm square with midpoint pitch and corner roll. Contact tasks hold a 15 N processed-compression setpoint during axial roll or during ±15° and ±30° pitch. Swept volume is reconstructed from recorded proximal trajectories on the arm collision model. It excludes the initial occupancy and is not a direct volume measurement.

Spatial and measured-load statistics, mean ± sample SD, n = 3. Swept volume uses the executed interval. Peak torque, raw RMS, and excursion use motion samples only.
Task	Wrist	Volume (L)	Peak (N·m)	Raw RMS (N·m)	Excursion (N·m)
Free axial	Locked	27.28 ± 0.03	53.10 ± 0.07	18.932 ± 0.005	23.226 ± 0.027
Free axial	Active	1.25 ± 0.01	30.24 ± 0.08	14.322 ± 0.015	2.454 ± 0.052
Free square	Locked	70.84 ± 0.04	65.89 ± 0.10	19.426 ± 0.009	21.816 ± 0.080
Free square	Active	30.03 ± 0.09	56.22 ± 0.04	19.159 ± 0.022	11.691 ± 0.145
Contact roll	Locked	26.80 ± 0.05	48.94 ± 0.99	16.362 ± 0.012	22.491 ± 0.044
Contact roll	Active	1.13 ± 0.01	25.27 ± 0.09	11.484 ± 0.007	2.729 ± 0.057
Contact pitch ±15°	Locked	14.19 ± 0.02	57.63 ± 0.43	21.513 ± 0.007	5.490 ± 0.210
Contact pitch ±15°	Active	7.85 ± 0.01	55.54 ± 0.19	21.490 ± 0.011	4.450 ± 0.024
Contact pitch ±30°	Locked	27.78 ± 0.04	60.88 ± 0.66	21.651 ± 0.006	9.948 ± 0.039
Contact pitch ±30°	Active†	15.41 ± 0.01	58.90 ± 0.31	21.781 ± 0.007	7.364 ± 0.108

† Active ±30° contact pitch includes internal pauses. Pauses stay in the swept-volume interval and are excluded from the motion-only torque columns.

Measured proximal torque for contact roll and the square task, plus per-joint peak torque. — Measured proximal torque. Raw joint-2 traces for contact roll and the square task are from repeat 1, without smoothing or time warping. The bar chart is per-joint contact-roll peak torque, mean ± sample SD, n = 3. Joint 2 carries the largest absolute load and does not represent all seven joints.

Contact-roll joint torque, peak torque, force regulation, and orientation tracking for locked and active wrists. — Contact roll in more detail. Peak proximal torque falls from 48.94 ± 0.99 N·m to 25.27 ± 0.09 N·m. Processed compression stays near the 15 N setpoint, while forward-kinematics orientation error is larger with the wrist active.

Task fidelity

The spatial and load reductions come with a tracking cost. Orientation error from arm–wrist forward kinematics was larger for the active wrist in every condition. Position error is the free-space forward-kinematics error, or the tangential error during contact. Force error is a processed flange-sensor estimate, not an independent measurement. The ±30° contact-pitch trials also contained 2.43 s of internal pauses.

Motion-window means over three trials. These are not peak errors or independently measured absolute accuracies.
Task	Wrist	Position (mm)	Orientation (°)	Force (N)	Pause (s)
Free axial	Locked	0.635	1.469	—	0.00
Free axial	Active	0.536	3.264	—	0.00
Free square	Locked	0.542	0.341	—	0.00
Free square	Active	1.259	1.423	—	0.00
Contact roll	Locked	0.354	0.657	0.176	0.00
Contact roll	Active	0.611	1.614	0.126	0.00
Contact pitch 15°	Locked	0.432	0.261	0.209	0.00
Contact pitch 15°	Active	1.108	0.729	0.206	0.00
Contact pitch 30°	Locked	0.614	0.476	0.605	0.00
Contact pitch 30°	Active†	1.596	1.182	1.059	2.43

Discussion

Distal redundancy moves reorientation toward the tool and reduces proximal motion. In the square task, proximal travel stayed similar during translation and fell during pitch and especially roll. Free-space axial rotation and contact roll also reduced measured peak and raw-RMS proximal torque. Pitch and the composite square kept the swept-volume and torque-excursion benefits with little change in raw RMS. A smaller motion envelope therefore does not, by itself, imply a comparable drop in raw joint torque.

The add-on wrist works with the arm’s existing Cartesian motion–force controller. Contact trials preserved force regulation. Larger active orientation error points to coordinated tracking and wrist–tool mounting as the next design constraints. The same-setting repetitions do not establish cross-workspace generalization, whole-system energy savings, or clinical imaging quality.

Citation

This manuscript has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible.

@misc{ren2026distalshift,
  title = {DistalShift: How Distal Redundancy Shapes Robot Arm Motion and Joint Loading},
  author = {Ren, Tian-Ao and Wang, Tongyu and Zhao, Shixuan and Karunaratne, Rohan and Zhang, Tao and Thomasson, Rachel and Cutkosky, Mark},
  year = {2026},
  note = {Manuscript submitted to the IEEE for possible publication},
  url = {https://ruggedizedwrist.github.io/}
}

Download PDF