Source-linked AI summary
A Controlled Comparison of Manual and Teleoperated Intraocular Instrument Motion for an Input Device
Korab Hoxha, Mirza Imamovic, Angelo Henriques, M. Ali Nasseri
TL;DR
The claim that surgical input devices preserve trained technique is rarely measured, and prior comparisons could not hold the physical instrument fixed. This study compares manual and teleoperated intraocular motion using a shared RCM frame; task outcomes were equal, but execution differed across measures.
Problem
Input devices are often said to preserve trained movement, but direct evidence is limited, and prior comparisons changed the instrument between manual and robotic conditions.
Source: §1 · p. 2, §2.2 · p. 4Method
The paper recovers the RCM from tracked instrument trajectories to compare manual and teleoperated motion in a common frame.
Source: §7 · p. 15Results
Teleoperated trials took three times as long, ran at a quarter of the speed, covered under 40% of the angular territory, and had 3.5 times as many movements, while task outcome was equal but at ceiling.
Source: §7 · p. 15Takeaways & Limitations
Reducing instrument degrees of freedom did not reduce manual effort: finger activity doubled and pinch variability tripled as effort shifted into repeated press-release cycles.
Source: §7 · p. 15Instrumented gloves perturbed performance and were used only in a terminal trial, confounding their effects with fatigue and order; tip tremor was unavailable in either condition.
Source: §6.6 · p. 15
Abstract
from arXiv · showhide
Input devices for robotic microsurgery are frequently described as preserving the surgeon's trained technique, but the claim is rarely measured. We compared manual and teleoperated intraocular instrument motion with the trocar constraint, the instrument, the eye model and the tracking source common to both conditions, so that the control interface was the only factor varied. Prior comparisons cannot hold the instrument fixed, because a robotic instrument is not the tool used manually. Sixteen participants performed a navigation task on a commercial ophthalmic simulator by hand and through a three-degree-of-freedom input device commanding a five-joint robot. Task outcome was equal but at ceiling: every participant acquired all five targets under both interfaces with no retinal or lens injury. Execution differed on every measure. Teleoperated trials took three times as long at a quarter of the median speed, covered less than half the angular working range, and were broken into 3.5 times as many separate movements. Completion time and movement fragmentation improved substantially across four trials of practice and had not plateaued; the measures set by the configured rate ceiling and joint limit changed the least. Finger activity doubled and pinch variability tripled, so reducing instrument degrees of freedom redistributed manual effort rather than reducing it. The interface preserves the outcome and reshapes the execution.
1 Introduction
Vitreoretinal surgery demands precise motion around fragile anatomy, while the claim that surgical input devices preserve trained movement is rarely tested. This work compares manual and teleoperated motion under shared instrument, trocar, eye-model, and tracking conditions.
1 Introduction
- The retina and vitreoretinal structures are fragile, while hand tremor can exceed the scale of structures that require precise positioning.The internal limiting membrane is 2–3 µm thick, and physiological tremor spans 100–200 µm peak-to-peak.Source: §1 · pp. 1–2
- Input devices are often described as preserving natural surgical technique without measuring whether they elicit the same motion as manual operation.A direct comparison of the same surgeon’s motion by hand and through an interface on the same task would make that claim testable.Source: §1 · p. 2
- Prior comparisons confound interface effects with changes in the instrument or kinematic constraint, whereas intraocular surgery permits the same instrument and trocar constraint in both conditions.The eye model and tracking source can also be held common, isolating the teleoperation pathway and control architecture.Source: §1 · p. 2
- Sixteen participants performed a simulator navigation task manually and through EMIRAS, a three-input interface commanding a five-joint robot.The study examines task success and movement behavior, including speed, path length, angular workspace, fragmentation, mode engagement, practice effects, and finger kinematics.Source: §1 · p. 2
- The work contributes a per-trial RCM frame for comparing manual and teleoperated motion, a controlled comparison, and quantitative measures of clutching, tremor, and learning.The RCM is recovered from the instrument trajectory and motion is decomposed into insertion depth and aiming angles; the contribution is its shared application, not a new method.Source: §1 · p. 3
2 Related Work
Prior work characterizes surgical input devices by specifications and task outcomes more often than by the motion they elicit. This section identifies gaps in controlled kinematic comparison, clutching, tremor measurement, and finger-level measurement that motivate the present work.
2.1 Input Devices for Robotic Microsurgery
- Commercial and application-specific interfaces span general-purpose and ophthalmic designs, but are typically characterized by specifications rather than correspondence with manual movement.Ophthalmic devices include four-DOF controllers with motion scaling and tremor filtering, and a six-DOF approach that constrains arm posture.Source: §2.1 · p. 3
2.2 Kinematic Comparison of Manual and Teleoperated Operation
- Prior manual-versus-teleoperation studies report measurable movement differences, but comparisons change the tool and sometimes the kinematic constraint, and equivalent intraocular evidence is not known.The cited comparison literature includes laparoscopic and general-surgery work on movement, posture, coordination, and task outcomes.Source: §2.2 · pp. 3–4
2.3 Workspace, Scaling and Clutching
- Workspace and clutching have been studied, but engagement behavior as a consequence of a specific surgical input-device design has not been reported in the cited literature.The gap concerns duty cycle, engagement duration, and engagement count alongside the workspace and mode structure.Source: §2.3 · p. 4
2.4 Tremor Measurement
- Prior tremor studies generally measure a single point; measurement across a teleoperation chain in a common unit, with hand tremor separately measured, is unreported in the reviewed work.The paper notes that the measured answer can differ qualitatively depending on where in the chain tremor is measured.Source: §2.4 · p. 4
2.5 Hand-Level Measurement in Surgery
- Finger-level kinematic measurement in ophthalmic microsurgery was rare, and the prior example cited here assessed manual technique only.The cited work used stretchable strain sensors on all fingers during simulated intrascleral intraocular lens fixation.Source: §2.5 · p. 4
2.6 Position of This Work
- This work compares motion elicited by an input device with manual technique at both hand and instrument levels, adding clutch statistics and tremor projected from handle to tip.Figure 2 shows the device’s two gimbal rotations and lever input, with a button selecting aiming or insertion and corresponding robot rotations about the trocar.Source: §2.6 · p. 4, Figure 2 · p. 5
3 System Description
EMIRAS uses a pen-grasp stylus and a phase-specific mapping to control the robot, with a mechanical arrangement intended to separate aiming from insertion. Its grip-to-tip displacement relation varies with insertion depth.
3.1 Design Rationale
- EMIRAS uses a pen grasp with the wrist rested and a distal pivot, matching the posture and fulcrum relationship of a trocar-constrained manual instrument.Three sensed inputs are multiplexed onto five robot joints through phase-specific modes rather than mapped simultaneously.Source: §3.1 · p. 5
3.2 Mechanical Design and Instrumentation
- The stylus has a two-axis gimbal, a spring-loaded lever input, and a momentary button that selects between rotation and insertion.The gimbal axes intersect at pivot P; the operator grips 100 mm from P, and lever deflection is kinematically independent of the gimbal.Source: §3.2 · p. 5
3.3 Input Variables and Scaling Geometry
- The input vector comprises two gimbal angles and lever deflection; the lever does not displace the grip point, separating aiming from insertion.The grip and instrument tip move on arcs about their respective pivots, so the effective hand-to-tip displacement gain varies with instrument depth.Source: §3.3 · pp. 5–6
3.4 Robot
The robot is modeled as a reduced five-joint chain, while its interface maps handle motion into trocar-centered aiming and insertion. Separate modes, frame alignment, and bounded velocity commands govern setup and intraocular motion.
3.4 Robot
- The robot reduces to three translational and two rotational joints, and holds the same unmodified Eyesi instrument used manually.The reduced chain is the model used throughout the work.Source: §3.4 · p. 6
3.5.1 Frame alignment
- A fixed 300° frame rotation aligns device angles to the robot because its rotational joints are not parallel to the microscope view.The alignment preserves correspondence between hand direction and visually perceived direction.Source: §3.5.1 · p. 7
3.5.2 Motion scaling
- With g = 3.0, ±24° gimbal deflection spans the robot’s full ±8° rotational range.The scaling maps hand-scale excursions to sub-tremor-amplitude motion.Source: §3.5.2 · p. 7
3.5.3 Operating modes
- Free mode positions the instrument at the trocar during setup; inside the eye, button-held RCM aims and released Depth inserts with rotations held.Free mode is not exercised in recorded trials.Source: §3.5.3 · p. 7, Figure 3 · p. 7
3.5.4 Rotational law
- The rotational law applies proportional velocity to tracking error, with G = 4.0 s−1, V = 1.5°/s, a 0.1° deadband, and a slew limit.Full command is reached at 0.375° of error.Source: §3.5.4 · pp. 7–8
3.5.5 Insertion law
- Lever deflection commands axial velocity through a deadband, capped at 1.5 mm/s and bounded to −12/+15 mm about home.The insertion command is governed by the deadbanded lever input.Source: §3.5.5 · p. 8
3.6 Timing and Response
End-to-end latency was measured from handle motion to the robot’s first movement using an optical camera. The measured latency was below 10 ms.
3.6 Timing and Response
- < 10 ms was the measured latency from handle motion onset to the robot’s first movement.Latency was measured optically with a 240 Hz camera over 50 trials.Source: §3.6 · p. 8
- Latency timing counted frames from input onset to first robot movement.The optical measurement used a 240 Hz camera.Source: §3.6 · p. 8
- 50 trials were used to measure end-to-end latency with a 240 Hz camera.The timing window ran from handle motion to first robot movement.Source: §3.6 · p. 8
3.7 Feedback
The interface provides motor-based workspace feedback, but force rendering was disabled in this study. The host processes inputs and commands the robot while synchronized logs record device and control state.
3.7 Feedback
- Force rendering was disabled and no haptic condition was evaluated; the active trial feedback was a vibrotactile cue near the rotational joint limit.The repeating pulse fired within 0.5° of the ±8° axis limit.Source: §3.7 · p. 8
3.8 Workspace
- The grip point spans a spherical cap of radius 100 mm and ±40° per axis, sweeping the instrument ±8° about the trocar at g = 3.0.Source: §3.8 · p. 8
3.9 Software and Logging
- The host reads the input at ∼490 Hz, applies multiplexing, frame rotation, scaling, and control, then sends joint commands over a 100 Hz serial link.Source: §3.9 · p. 8
- Synchronized kinematics and control logs share a clock, and each session archives its configuration for trial-level parameter recovery.The kinematics stream runs at the uplink rate and the control stream at 100 Hz.Source: §3.9 · p. 8
4 Methods
Sixteen participants performed a five-target navigation task under manual and teleoperated conditions, with the same eye model, trocar, instrument, and tracking source. The study recorded simulator, device, and glove data, derived shared motion measures, and analyzed repeated trials with mixed models.
4.1 Participants
- The sample comprised 3 vitreoretinal surgeons, 6 residents, and 7 novices; experience was not analyzed as a factor.The relationship between experience and movement was reserved for future work.Source: §4.1 · p. 8
4.2 Task
- Participants acquired five of twelve retinal targets without injury instructions focused on speed, and no time limit was imposed.The task instruction was to acquire targets without injuring the retina.Source: §4.2 · p. 8
4.3 Conditions and Protocol
- The comparison held the eye model, trocar, instrument, and tracking system common across conditions, varying the control interface.Condition order was randomized and each condition began with a warm-up.Source: §4.3 · p. 9
4.4 Recorded Data
- Simulator trajectories and per-trial scores were supplemented by teleoperation input/control logs and glove recordings of 20 finger-joint angles.The simulator trajectory was exported at 30 Hz, and glove angles were recorded at 60 Hz.Source: §4.4 · p. 9
4.5 Measures
- Tool-tip motion was expressed in spherical coordinates about a per-trial trocar estimated by least-squares intersection, identically for both conditions.This common frame allowed motion comparison despite manual surgery having no mode structure for segmentation.Source: §4.5 · pp. 9–10
- Each trial’s motion window used a threshold set to 25% of that trial’s 90th-percentile speed to avoid biasing the slower teleoperated condition.Movement segments were contiguous intervals above 5% of the trial’s 90th-percentile speed for at least 0.4 s.Source: §4.5 · p. 10
- Teleoperation-only measures captured button duty cycle and engagements, while glove measures captured finger activity, pinch variability, and submovements.The glove measures were computed from 20 joint angles and four pinch measures.Source: §4.5 · p. 10
4.6 Analysis
- Linear mixed models tested condition, trial, and their interaction across four ungloved trials, with participant random intercepts; glove trials used signed-rank tests.Effect sizes and confidence intervals were primary evidence given the 16-participant sample.Source: §4.6 · p. 10
5 Results
Task success was at ceiling, but teleoperation substantially reshaped movement; practice narrowed some differences while configured hardware limits persisted. The results also show that low-rate path reconstruction is not comparable across conditions and that teleoperation increased measured finger activity.
5 Results
- 149 trials were analyzed: 122 ungloved trials entered mixed models, while 27 glove trials carried tool-tip data.Source: §5 · p. 10
5.1 Task Outcome
- All participants acquired all five targets in both conditions, with no retinal or lens injuries; completion was at ceiling and did not establish equal precision.The study was not designed or powered to test precision equivalence.Source: §5.1 · pp. 10–11
5.2 Measurement Validity
- Reconstructed path over-read the simulator odometer by 5% manually and 29% teleoperated; duration-correlated noise made path length incomparable across conditions.Filtering also failed to correct both conditions simultaneously, so path results use the odometer.Source: §5.2 · p. 11, Figure 5 · p. 11, Table 2 · p. 11
5.3 How the Task Was Performed
- 3× as long, at roughly a quarter of median speed, teleoperated trials covered less than half the angular territory and used 3.5× as many movements.The instrument path was about a quarter shorter under teleoperation.Source: Table 3 · p. 12, §5.3 · p. 12, Figure 6 · p. 12
5.4 Rate Ceilings
- The robot’s 1.5°/s rotational and 1.5 mm/s translational limits were below manual peak aiming speed of 14.4°/s; velocity limits were a major contributor to teleoperation’s temporal cost.A path ratio of 0.79 and mean-speed ratio of 0.25 predicted a 3.19 duration ratio, versus 2.98 observed.Source: §5.4 · p. 12
5.5 Clutching
- Aiming commands reached the robot during 38% of the session, and 29 press–release cycles occurred per trial, far more than the five to ten forced by mode requirements.Neither aiming nor insertion commands were active during 40% of trial time; the data do not distinguish what occurred in that interval.Source: §5.5 · pp. 12–13, Table 4 · p. 13
5.6 Practice
- Completion time improved by 27 s per trial and movement segments by 14 across four trials, with neither measure plateauing; other gaps changed least where hardware limits governed them.Teleoperated improvement was steep in both experience subgroups; the path interaction classification was uncertain.Source: §5.6 · p. 13, Figure 7 · p. 13
5.7 Hand Kinematics
- In glove trials, teleoperation approximately doubled finger activity (134.9 vs 71.2 summed joint standard deviations), pinch variability (10.0 vs 3.3), and submovements (115 vs 55).Finger tremor-band power did not fall; the glove results are limited by order and fatigue confounds, glove-induced slowing, one trial per condition, and uncharacterized frequency response. The projected tip tremor was 1.55 µm RMS, not a measurement, versus 38 µm RMS reported for unassisted surgery.Source: §5.7 · pp. 13–14
6 Discussion
Task success was equal, but teleoperation reshaped movement; the discussion attributes some costs to robot limits and others to control design, while identifying simulator and measurement boundaries.
6.1 What the Interface Preserves and What It Changes
- Teleoperation altered movement despite equal task success: motion was slower, more fragmented, and confined to a smaller angular range.The interface’s manual-like posture and aiming gestures did not ensure manual motor behavior once scaling, mode switching, workspace constraints, and robot dynamics intervened.Source: §6.1 · p. 14
- Aiming uses position mapping with a fixed 3:1 angular reduction, whereas insertion uses rate control rather than manual-distance mapping.The design premise of preserving manual movement is retained for aiming and abandoned for insertion.Source: §6.1 · p. 14
6.2 The Velocity Limit Is a Major Constraint
- The robot’s velocity limit, set by joint accuracy, accounts for the threefold completion-time difference with no residual.Manual peak aiming exceeded the robot’s limit by roughly tenfold; the measures governed by velocity and joint range also showed the smallest practice effects.Source: §6.2 · p. 14
6.3 Clutching
- Operators engaged the rotational command 29 times per trial; mode switching explains perhaps a third, but not all of those engagements.Allowing insertion without releasing aiming, or switching modes automatically on depth, could remove the forced component; its size remains unresolved.Source: §6.3 · p. 14
6.4 The Hand Works Harder
- Finger activity doubled and pinch variability tripled as work shifted into repeated press-release cycles, rather than decreasing with fewer instrument degrees of freedom.The direction of the finger-level difference is robust, but its magnitude is difficult to calibrate because comparable measurements are scarce.Source: §6.4 · p. 14
6.5 Design Implications
- The temporal cost requires a faster, still-accurate manipulator; roughly 40% of the session was spent in neither aiming nor inserting.A context-varying speed limit could bind near tissue yet relax during gross repositioning; reconsidering the multiplexed button targets time spent disengaged.Source: §6.5 · p. 15
6.6 Limitations
- The simulator omitted trocar friction, tissue compliance, and error consequences; one task excluded peeling, injection, and bimanual work, and the light pipe was fixed.Force rendering and workload ratings were absent; glove effects were confounded with fatigue and order, and neither condition had tip-tremor measurement.Source: §6.6 · p. 15
7 Conclusion
With the trocar constraint, instrument, eye model, and tracking source held common, the study compared manual and teleoperated motion. Outcomes matched at ceiling, but execution differed and the comparison also yielded a common motion-analysis approach.
7 Conclusion
- Teleoperated trials took three times as long at a quarter of the speed, covered under 40% of the angular territory, and used 3.5 times as many movements, despite equal task outcome at ceiling.Completion time and movement fragmentation improved substantially across four trials without plateauing; measures set by the robot’s rate limit did not.Source: §7 · p. 15
- The temporal cost follows from the robot’s accuracy-linked velocity limit, while engagement frequency and lever deadband follow from control design and could change without new hardware.Reducing instrument degrees of freedom did not reduce manual effort: finger activity doubled and pinch variability tripled as work shifted into press-release cycles.Source: §7 · p. 15
- Recovering the remote center of motion per trial and decomposing motion around it enables the same comparison for manual and teleoperated operation using a tracked instrument.Reconstructing path length from decimated position exports can inflate it with trial duration and nearly produced a false equivalence between conditions of unequal length.Source: §7 · p. 15
Conflict of Interest
The authors report that the input device and robot were developed by their group and disclose partial support from NSK Ltd.
Conflict of Interest
- The input device and the robot it commands were developed by the authors; the work received partial NSK Ltd. support, with no funder role reported.The authors declare no other competing interests.Source: Conflict of Interest · p. 15
Author Biographies
The biographies describe four authors’ engineering, medical-technology, robotics, and surgical-research backgrounds.
Author Biographies
- Korab Hoxha studied mechanical and medical technology engineering and is pursuing a Ph.D. developing a robotic system for eye surgery.His research interests include surgical robotics, biomechanics, and automation in medicine.Source: Author Biographies · p. 19
- Mirza Imamovic studied mechanical engineering and mechatronics and robotics, and works on teleoperation, haptic feedback, and human–robot interaction in surgical applications.He is a research assistant responsible for establishing and coordinating the Ergonomics Evaluation Lab.Source: Author Biographies · p. 19
- Angelo Henriques researches integrating scene graphs and multimodal sensor data into robotic surgical workflows for ophthalmic microsurgery.He is a doctoral candidate at TUM and a Research Associate at TUM University Hospital.Source: Author Biographies · p. 19
- M. Ali Nasseri is Professor of Surgical Robotics at TUM and founding director of the MAPS Laboratory at TUM University Hospital.His research focuses on surgical robotics, autonomous surgery, surgical intelligence, and AI-driven precision surgery.Source: Author Biographies · p. 19