Surgical Manipulation Interface

SuMI:Zero-Shot Policy Transfer between Different Robots with Wristed Instruments for Minimally Invasive Surgery

  • Nirmal Pol1,2,*
  • Radian Gondokaryono2,3,*
  • Puspita Triana Dewi4
  • Fata A. Niazi3
  • Jones Law
  • Kashfia Mahmood
  • Lueder A. Kahrs1,2,3,5

* Equal contribution

University of Toronto

  • 1Institute of Biomedical Engineering
  • 2Medical Computer Vision and Robotics Lab
  • 3Department of Mathematical and Computational Sciences
  • 4Continuum Robotics Lab
  • 5Department of Computer Science

Learn on a serial arm. Transfer to a surgical robot.
Matching the tool, observations, and constrained motion across embodiments.

Read the paper PDF GitHub repository Explore the results
See the RCM constraint in motion. Franka SuMI (left) and the dVRK (right) pivot their instruments about a fixed remote center of motion, matching the constraint at a surgical entry point. Original playback speed, without audio.
Franka SuMI experimental setup with a wristed EndoWrist tool, wrist camera, stereo endoscope, and angled peg transfer task.
01 / Source robotFranka SuMICollect demonstrations and train the policy.
da Vinci Research Kit experimental setup using the same EndoWrist instrument and camera arrangement for zero-shot policy deployment.
02 / Target robotda Vinci Research KitDeploy the same policy without fine-tuning.
Franka SuMI demonstrations
206
Successful zero-shot dVRK trials
17 / 20
Target-robot training demonstrations
0

Overview

A shared interface for surgical policy learning

Abstract

Learning-based surgical autonomy is constrained by limited access to specialized surgical robots such as the da Vinci Research Kit (dVRK). Serial manipulators are widely available and some equipped with surgical tools can support policy development, but have yet to be used for training policies that deploy on surgical robots. Surgical Manipulation Interface (SuMI) reduces embodiment gaps between serial arms and surgical robots to enable zero-shot policy transfer.

First, SuMI equips a serial manipulator with EndoWrist surgical instruments to achieve a matched visual appearance of the end effector and kinematics of the tool wrist joints. Second, we develop a controller with RCM constraint enforcement, 6D tool-tip control, and arm manipulability optimization to match the kinematically constrained trocar motion of the tool. Third, we develop a novel RCM-constrained manipulability metric to compare software RCM serial manipulators and mechanical RCM surgical robots on a common task-space velocity scale, and use it to show that our Franka serial arm with surgical tool (Franka SuMI) reproduces the task-relevant workspace and Cartesian execution speeds of the dVRK. Finally, we show that a Diffusion Policy trained on an angled peg transfer task collected on Franka SuMI is deployed on the dVRK, achieving 17/20 successful trials.

Matched surgical tools

A modular actuation unit brings EndoWrist instruments to serial robot arms, matching the surgical tool’s appearance and wrist motion.

RCM-constrained control

A hierarchical controller prioritizes the remote center of motion, then 6D tool-tip pose tracking, then arm manipulability.

Transfer across embodiments

A common metric compares tool-tip motion capabilities. A policy learned on Franka SuMI transfers directly to the dVRK.

Different robots. A shared interface.

Closing the embodiment gaps

SuMI aligns what the policy sees, how the instrument moves, and the motion each robot can execute.

Visual observations

The same instrument, wrist camera, and stereo endoscope views reduce differences in what the policy sees.

Matched tool & camera views

Tool interface

Coordinated roll, pitch, yaw, and jaw actuation reproduce the motions of the EndoWrist instrument.

Shared wristed instrument

Kinematic constraints

Software RCM on Franka SuMI reproduces the pivoting motion of the dVRK’s mechanical RCM.

Same entry-point constraint

Execution capabilities

RCM-constrained workspace analysis compares the tool-tip velocities available on both robots.

Common velocity scale
One learned policy.

The scene, cameras, and instrument move from Franka SuMI to the dVRK. The Diffusion Policy then runs without retraining or fine-tuning.

Hardware & control

A modular EndoWrist actuation unit

The 3D-printed assembly couples four servo motors to the instrument’s actuation dials. Parametric geometry adapts the mounting interface to a serial arm.

For Franka SuMI, the seven arm joints and three instrument joints form a unified 10-DoF kinematic chain, with separate jaw actuation.

  1. Enforce the RCM constraintKeep the instrument shaft passing through the desired entry point.
  2. Track the tool-tip poseControl Cartesian position and orientation within that constraint.
  3. Optimize arm manipulabilityUse the remaining redundancy to improve the arm configuration.

The implementation combines libfranka and ROS 2, with a 1 kHz arm control loop and a 400 Hz RCM controller in the reported experiments.

Exploded CAD view of the SuMI actuation unit, including the motor holder, servo motors, shafts, orthoplanar springs, coupling interfaces, and adjustable tool holder.
The modular actuation unit and its adjustable mounting geometry.

Experiments

Angled peg transfer across embodiments

A Diffusion Policy trained on 206 demonstrations from Franka SuMI successfully completed 85% of trials on the dVRK, with no target-robot demonstrations, retraining, or fine-tuning.

Angled peg transfer performance
Evaluation conditionSuccessful trialsPick-up timeDrop-off time
Franka SuMI Source platform19 / 20 95%22 ± 5 s40 ± 8 s
dVRK Zero-shot transfer17 / 20 85%33 ± 10 s58 ± 14 s
Franka SuMI Dynamic RCM17 / 20 85%23 ± 10 s39 ± 12 s

Times are measured from the start of the rollout and reported as mean ± standard deviation over successful trials. Each condition includes 20 trials.

The same policy on two robots. Franka SuMI (left) and the dVRK (right) perform angled peg transfer. The policy transfers to the dVRK without retraining or fine-tuning. Shown at 1× speed, without audio.
Sequential wrist-camera and stereo-endoscope frames showing the same policy grasping, reorienting, and placing an angled peg on Franka SuMI and the dVRK.
The same learned policy on Franka SuMI and the dVRK. Highlighted frames show tool reorientation during grasping and placement. Select a figure to view it at full size.

Adapting the entry point

Franka SuMI also completed 17/20 trials while the desired RCM height varied sinusoidally to simulate breathing-induced motion. This motion was absent from the training demonstrations.

What the evaluation covers

The study evaluates one task and one source–target robot pair. Transfer required manual scene and camera alignment. Residual visual and geometric differences led to corrective motions and longer rollouts on the dVRK.

Dynamic RCM in motion. Franka SuMI performs angled peg transfer while the desired entry point moves vertically to simulate breathing. The red marker indicates the RCM. Shown at 1× speed, without audio.
Explore the RCM-constrained workspace analysis
Franka SuMI workspace plots showing RCM-constrained minimum tool-tip velocity and condition number within a pyramidal surgical workspace.
RCM-constrained tool-tip velocity capability and condition number.

Comparing motion on a common scale

The RCM-constrained, scaled Jacobian accounts for the trocar constraint and joint velocity limits. Its singular values characterize tool-tip motion capabilities across different robot embodiments.

The smallest reported minimum singular value was 0.068 m/s throughout the evaluated Franka SuMI workspace, exceeding the 0.028–0.030 m/s operational benchmark used in the paper.

This comparison concerns operational tool-tip speeds in the evaluated workspace. It does not establish that Franka SuMI matches the dVRK’s absolute mechanical speed limit.

Resources

Explore SuMI

Read the paper for the complete method and experiments. Code and design files will be added to the project repository.

Research paperFull PDF · 8.8 MB Project repositoryMEDCVR / SuMI on GitHub

Citation

Cite SuMI

If you use SuMI in your research, please cite our work using the entry below.

BibTeX 
@misc{pol2026sumi,
  author = {Pol, Nirmal and Gondokaryono, Radian and
            Dewi, Puspita Triana and Niazi, Fata A. and
            Law, Jones and Mahmood, Kashfia and Kahrs, Lueder A.},
  title  = {{SuMI}: Zero-Shot Policy Transfer between Different Robots
            with Wristed Instruments for Minimally Invasive Surgery},
  year   = {2026},
  url    = {https://medcvr.utm.utoronto.ca/SuMI.html}
}