FOMIA Force-Feedback Manipulation Interface for
Teaching Contact-Rich Manipulation

Yishu Shen1Yiming Luo2†Ruilin Wang3Bowen Yang3Shuai Yang1Fangyi Xu3Yujun Shen3Fu Zhang2Boyu Zhou4Yinghao Xu1*
1 The Hong Kong University of Science and Technology2 The University of Hong Kong3 Robbyant4 Southern University of Science and Technology

† Project leader   ·   * Corresponding author

Paper BibTeX
Scroll to explore

OVERVIEW

Overview of FOMI. The force-feedback manipulation interface incorporates a scissor-shaped mechanism to provide human operators with intuitive force feedback in robot-free collection of contact-rich demonstrations, while also supporting teleoperation for policy learning.

Overview of FOMI: intuitive force feedback in robot-free collection, teleoperation, demonstration collection performance, and contact-rich policy learning.

Mechanical Design For Intuitive Force Feedback

Actuation links in FOMI’s scissor-shaped mechanism transmit object interaction forces directly to the handles without force sensors, providing intuitive force feedback, which supports real-time adjustment in contact-rich tasks. Meshed sector gears synchronize gripper opening and closing, while the handles support the interface during manipulation.

Mechanical design of FOMI: exploded view, direct force transmission to the handles in robot-free collection, kinematic schematic, opening and closing motion, and mechanical stops.
Direct force transmission path from the gripper to the handles in robot-free collection, and mechanical stops defining the gripper width limits. At full opening, the maximum gripper width reaches 84.9 mm.
Sensor layout Click to expandClick to collapse
Sensor layout on FOMI and the robot gripper: a 6-DoF pose tracker, fisheye RGB cameras, and magnetic rotary encoders.
Sensor layout of FOMI and the robot gripper.

Demonstration Collection

FOMI supports robot-free collection through sensor calibration, data acquisition, and synchronization, and enables teleoperation by mapping pose and gripper measurements to robot commands. Teleoperation relies on visual feedback; mechanical stops indicate gripper width command limits without conveying object interaction forces.

Robot-free collection

Speed: 1×

Teleoperation

Speed: 2×

Contact-Rich Manipulation Tasks

Representative rollouts of policies trained on robot-free demonstrations across four tasks.

Speed: 1×

Deployment across robot embodiments

Speed: 1×

Results

Demonstration Collection Performance

In an exploratory study, three operators with no prior robot operation experience used each interface on all four tasks. FOMI achieved the highest mean collection efficiency on each task and the highest overall completion rate in both robot-free collection and teleoperation.

Robot-free collection Click to expandClick to collapse
Robot-free collection interfaces: FOMI, UMI, and RING.
Successful demonstrations / 10 min
    Robot-free collection efficiency, mean successful demonstrations per 10-minute session
    TaskFOMIUMIRING
    Pick Chip22.016.610.7
    Insert Tube17.713.89.9
    Wash Dish13.311.010.9
    Open Pouch18.213.715.0
    Teleoperation Click to expandClick to collapse
    Teleoperation interfaces: FOMI, GELLO, and VR.
    Successful demonstrations / 10 min
      Teleoperation collection efficiency, mean successful demonstrations per 10-minute session
      TaskFOMIGELLOVR
      Pick Chip7.26.40.2
      Insert Tube7.44.91.6
      Wash Dish9.38.87.2
      Open Pouch9.78.14.6

      Policy Performance

      We fine-tune the π0.5 vision-language-action (VLA) model separately on robot-free and teleoperated demonstrations. The policies achieve average success rates of 78.4% with 100 robot-free demonstrations per task and 70.5% with 25 teleoperated demonstrations per task.

      Robot-free demonstrations

      Policy performance with 25, 50, and 100 robot-free demonstrations per task: success rate and progress rate.
      Policy performance with increasing numbers of robot-free demonstrations per task: 25, 50, and 100. Left: success rate. Right: progress rate.

      Teleoperated demonstrations

      Policy performance with 25 teleoperated demonstrations per task: success rate and progress rate.
      Policy performance with 25 teleoperated demonstrations per task. Left: success rate. Right: progress rate.

      Cite FOMI

      @misc{shen2026fomi,
  title = {FOMI: A Force-Feedback Manipulation Interface for Teaching Contact-Rich Manipulation},
  author = {Yishu Shen and Yiming Luo and Ruilin Wang and Bowen Yang and Shuai Yang and Fangyi Xu and Yujun Shen and Fu Zhang and Boyu Zhou and Yinghao Xu},
  year = {2026},
  eprint = {2610.19839},
  archivePrefix = {arXiv},
  primaryClass = {cs.RO},
  url = {https://arxiv.org/abs/2610.19839},
}