Overview

What is TRumi?

TRumi (Trossen Robotics Universal Manipulation Interface) is a handheld manipulation data collection system. It is a handheld parallel-jaw gripper with a wrist-mounted GoPro that records what the gripper sees and how it moves.

Instead of teleoperating a robot to gather demonstrations, an operator performs the task directly with the handheld gripper. This lets teams collect more manipulation demonstrations — across more objects, environments, and task variations — without setting up a full robot rig for every demo.

TRumi follows the Universal Manipulation Interface (UMI) approach: it captures demonstrations in end-effector space, rather than being tied to one specific robot’s joint configuration. This makes the collected data well suited to large-scale data collection across many environments, before any robot-specific training and validation.

Why collect data this way?

Intuitive. Operators perform tasks directly with the gripper, instead of reasoning about a robot’s joint configuration during teleoperation.
Fast, portable, anywhere. Scene setup is minimal — no robot, no fixed workcell. Collect in the environments where the task actually happens.
Scalable. Supports single-gripper and bimanual (two-gripper) collection, as well as multi-site collection across many operators.
Not tied to one robot. Because demonstrations are captured in end-effector space, the data is not tied to a specific robot’s joint configuration.

Things to keep in mind

TRumi is a complementary data collection tool, not a replacement for leader-follower teleoperation. Use TRumi to scale and diversify data collection; use leader-follower teleoperation for final tuning and validation on a specific robot.

Because TRumi captures data in end-effector space, it is not directly tied to any robot hardware. Deploying a trained policy on a real robot still requires integration with the target robot — converting policy outputs into robot motion, typically through interpolation, inverse kinematics, and validation on the actual hardware.

How it works

At its core, TRumi answers a single question: where was the gripper, and how open was it, at every moment of a demonstration? Because the GoPro is rigidly mounted to the gripper, tracking the camera through space is equivalent to tracking the gripper itself.

The workflow has three phases:

        flowchart LR
    A["<b>1 · Collect</b><br/>Perform the task with TRumi grippers<br/>GoPro records video + IMU data"]
    B["<b>2 · Process</b><br/>Visual-inertial SLAM estimates motion<br/>Extract pose and gripper width"]
    C["<b>3 · Output</b><br/>Package into a .zarr / .mcap dataset<br/>Use for downstream policy training"]
    A --> B --> C

    classDef box fill:#f2f2f2,stroke:#1e1d22,color:#1e1d22;
    class A,B,C box;

1. Collect. An operator performs the task while holding the TRumi gripper. The GoPro records the scene along with its onboard IMU (accelerometer and gyroscope) data. A one-time mapping video of the workspace is also recorded, giving SLAM a map to localize against.

2. Process. The dataset generation pipeline runs visual-inertial SLAM (ORB-SLAM3) to estimate the camera’s 6-DoF trajectory through the mapped scene, combining the video with the IMU data for robust, metric-scale motion. For each frame it then extracts the end-effector pose and the gripper width (how open the fingers are, measured from the finger identifiers).

3. Output. The synchronized frames, poses, and gripper widths are packaged into a structured .zarr or .mcap dataset, ready to feed into downstream policy-training workflows.

The rest of these docs cover each part of this workflow in detail. If this is your first time, continue to Specifications and then work through the setup pages in order.